Remote sensing image classification training sample optimization method based on ensemble learning

Abstract

The invention discloses a remote sensing image classification training sample optimization method based on ensemble learning, and relates to the technical field of remote sensing. Reserving areas with completely consistent classification results as reserved areas; completing sample quality optimization processing; constructing a sample point by using the geometric center of the plaque of the reserved area, and mixing the sample point with the optimized sample to realize sample number multiplication and layout optimization; on the basis of all unsupervised classification results and optimized sample overlay analysis, different sub-category samples in the same category are identified, and sample type refinement processing is completed; and comprehensively checking whether the sample purification and optimization treatment is effectively completed or not based on the comparison of visual interpretation classification results, the class separation degree change between the samples and the change of comparison precision evaluation. According to the method, the problems of uneven quality and uneven distribution of traditional remote sensing samples are effectively solved through a multi-classifier collaborative supervision classification and grid superposition analysis technology.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, and in particular to a remote sensing image classification training sample optimization method based on ensemble learning. Background Art

[0002] The continuously developing remote sensing technology is widely used in land use change research, land space planning, resource survey and natural disaster assessment. Remote sensing image classification is an important way to obtain remote sensing data information. It is also one of the important fields of remote sensing technology application research. It is also the basis and prerequisite for other statistical analyses. The accuracy of the classification results directly affects the development and application of remote sensing technology.

[0003] However, in the application of remote sensing technology, the complex natural environment causes phenomena such as "ground objects with similar spectral characteristics may have different categories, and different ground objects may also have similar spectral characteristics". Remote sensing image data may have problems such as noise, distortion and occlusion during the acquisition process. The singleness and limitations of image processing methods and the uncertainty of remote sensing image data classification results are inevitable. These factors all have an impact on the image classification results.

[0004] Traditional image classification methods are mainly divided into supervised classification and unsupervised classification. Supervised classification constructs a discriminant function based on training samples provided by the training area. The discriminant function classifies non-sample pixels into corresponding categories. Therefore, the quality of the training samples greatly affects the quality of supervised classification results. Many factors affect the quality of training samples, among which the quality, quantity, spatial distribution, and classification level of the training samples are the main factors affecting the accuracy of the classification results.

[0005] Although many mature image classification algorithms have emerged, the accuracy of their classification results is still uncertain and difficult to guarantee when the quality of training samples is low. Therefore, based on the idea of ​​classifier ensemble, four optimization processing methods are proposed to address four problems that arise with samples: sample quality screening, sample quantity expansion, sample layout optimization, and further refinement after optimization. Although the uncertainty of remote sensing image classification results cannot be completely eliminated, optimizing training samples through these four methods can improve the accuracy of remote sensing image classification and enable remote sensing technology to be better applied in various fields.

[0006] Currently, no effective solutions have been proposed for the problems in related technologies. Summary of the Invention

[0007] In response to the problems in the related art, the present invention proposes a remote sensing image classification training sample optimization method based on ensemble learning to overcome the above technical problems existing in the existing related art.

[0008] To this end, the specific technical solutions adopted in the present invention are as follows: The remote sensing image classification training sample optimization method based on ensemble learning includes: Initial samples were selected based on visual interpretation of land cover types, and an odd-numbered classifier was selected in the remote sensing image processing platform for supervised classification, and different land feature types were extracted. The classifier integration method was used to perform raster overlay on areas with different land feature types, and areas with completely consistent classification results were retained as retained areas. The initial training samples were superimposed with the retained areas, and samples outside the retained areas were screened and removed to complete the sample quality optimization processing. Sample points were constructed using the geometric centers of patches in the retained areas and mixed with optimized samples to achieve sample doubling and layout optimization. Sample type refinement was completed based on the overlay analysis of all unsupervised classification results and optimized samples, and different subcategories of samples in the same category were identified. Based on the comparison of visual interpretation classification results, changes in class separation between samples, and changes in comparative accuracy evaluation, a comprehensive test was conducted to verify whether sample purification and optimization processing were effectively completed.

[0009] Furthermore, the classifier integration method is used to perform raster overlay on areas of different land feature types, and the areas with completely consistent classification results are retained as the retained areas; the initial training samples are superimposed on the retained areas, and the samples outside the retained areas are screened and eliminated to complete the sample quality optimization processing, including: importing several classification results into the user desktop component, using the raster calculator to perform pixel-by-pixel overlay analysis on similar land features, and counting the pixel classification consistency; according to the pixel classification consistency results, they are divided into all inconsistent, partially consistent and all consistent, and all consistent areas are extracted as the retained areas; the initial training samples are spatially overlaid with the retained areas to determine whether the samples are located within the high-confidence retained area; if the sample is within the retained area and is of the same type as the regional land feature, the sample is retained, otherwise, the samples outside the retained area are screened and only the consistent samples within the area are retained to optimize the sample quality.

[0010] Furthermore, the geometric centers of the patches in the retained area are used to construct sample points and mix them with the optimized samples to achieve sample doubling and layout optimization, including: generating a grid of a preset format in the algorithm test area to achieve spatial partitioning, and counting the number of samples in each grid to identify dense and sparse sample areas; extracting the feature vectors of each sample in the dense area, and calculating the similarity between samples based on Euclidean distance, removing highly repeated samples, and retaining representative sample points; clustering the dense area samples using a spatial clustering algorithm, eliminating outliers and abnormal samples, and retaining a set of core sample points; extracting the geometric centers of the patches in the high-confidence retained area, generating sample points and merging them with the optimized dense area samples to achieve sample doubling; using an unbalanced learning mechanism in the sparse area to expand the number of samples in the sparse area.

[0011] The beneficial effects of the present invention are: 1. This invention effectively addresses the uneven quality and distribution of traditional remote sensing samples through the collaborative supervision of multiple classifiers and raster overlay analysis. A classifier integration strategy is used to extract high-confidence regions, significantly improving sample purity. Combined with patch geometry analysis and spatial block optimization techniques, the method achieves expanded sample size and balanced spatial distribution, enhancing the generalization and stability of the classification model. The introduction of unsupervised classification and hierarchical clustering further refines sample classification, thereby improving the accuracy of complex feature recognition.

[0012] 2. This invention combines ensemble learning with spatial analysis techniques to construct a comprehensive sample optimization technology system. Dynamic weight adjustment and adaptive screening mechanisms achieve intelligent improvement in sample quality. Optimization strategies based on geometric features and spatial distribution ensure the representativeness and balance of sample data. This entire process forms a closed-loop optimization system, significantly improving the automation and accuracy of remote sensing image classification, and providing reliable technical support for applications such as land use monitoring and environmental assessment.

[0013] 3. By introducing a dual-channel convolutional neural network and a dynamic quality assessment mechanism, this paper achieves a technological breakthrough in sample screening, transitioning from static rules to intelligent adaptation. The designed network architecture simultaneously extracts spectral features and spatial context information. Combined with adaptive threshold adjustment, it significantly improves the accuracy of distinguishing samples from complex feature boundaries. The two-stage application of transfer learning technology effectively addresses the model generalization problem in small sample areas, significantly improving classification accuracy and enhancing the cross-regional applicability of the method.

[0014] 4. This invention achieves balanced optimization of sample distribution through a spatial partitioning strategy, adopting differentiated processing schemes for different density areas, effectively improving the problem of uneven spatial distribution of samples. The collaborative analysis of unsupervised and supervised classification results makes the sample classification more refined, better reflecting the diversity characteristics of actual ground objects. This not only significantly improves the quality of sample data, but also lays a solid foundation for improving the accuracy of subsequent remote sensing image classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 is a flowchart of a remote sensing image classification training sample optimization method based on ensemble learning according to an embodiment of the present invention; Figure 2 2. It is a schematic diagram of a dual-channel CNN network design according to an embodiment of the present invention; Figure 3 Schematic diagram of a technical route for a remote sensing image classification training sample optimization method based on ensemble learning according to an embodiment of the present invention; Figure 4 is a schematic diagram of the number of sample points of various features in a sample set according to an embodiment of the present invention; Figure 5 is a comparison diagram of the initial sample distribution and the distribution of sample set 3 according to an embodiment of the present invention; Figure 6 is a distribution diagram of the purified sample according to an embodiment of the present invention; Figure 7 is a schematic diagram of the overall classification accuracy of multi-sample set training samples according to an embodiment of the present invention; Figure 8 2 is a schematic diagram of Kappa coefficients of multi-sample set training samples according to an embodiment of the present invention; Figure 9 2. This is a diagram showing the classification results of a CLCD dataset according to an embodiment of the present invention; Figure 10 is a sample classification result diagram according to an embodiment of the present invention; Figure 11 is a classified image before fine processing according to an embodiment of the present invention; Figure 12 This is a classified image after fine processing according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention. They are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. By referring to these contents, ordinary technicians in this field should be able to understand other possible implementation methods and the advantages of the present invention.

[0018] According to an embodiment of the present invention, a remote sensing image classification training sample optimization method based on ensemble learning is provided.

[0019] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, the remote sensing image classification training sample optimization method based on ensemble learning according to an embodiment of the present invention includes: S1. Select initial samples based on visual interpretation of land cover types, select an odd number of classifiers for supervised classification in the remote sensing image processing platform, and extract different land feature types; Specifically, all LUCC data were resampled to a uniform resolution and spatial reference. The LUCC classification system was unified based on the actual classification of features and the standard land use classification system. The intersection of multiple LUCC data sources was taken, retaining pixels that were consistent across at least N datasets to generate a high-confidence master sample library.

[0020] S2. Use the classifier ensemble method to perform grid overlay on regions of different land feature types, retaining regions with completely consistent classification results as retained regions; overlay the initial training samples with the retained regions, and filter out samples outside the retained regions to complete sample quality optimization; Specifically, select an odd number of classifiers (e.g., three) in ENVI for supervised classification, including a support vector machine (SVM), a maximum likelihood classifier (MAX), and a Mahalanobis distance classifier (MLD). Select a subset of the initial samples from the manual visual interpretation as training samples to ensure sufficient representative samples for each feature type. Set classifier parameters, such as the kernel function type and penalty parameter, based on the image's spectral characteristics and feature type. Perform supervised classification on the remote sensing image, generating multiple classification results.

[0021] Use the Raster Calculator in ArcMap to process multiple classification results: Use the Raster Calculator to write an expression to extract raster data of a specific feature type. For example, for water body type, use the expression Con("ClassifiedImage1"==1, 1) to extract water body raster data. Repeat the above operation for each classification result to extract raster data of different feature types such as water body, building land, cultivated land, bare soil, etc. Use the "Raster Overlay Analysis" tool in ArcGIS to overlay the raster data of the same feature type of multiple classification results. Count the consistency of feature types in multiple classification results pixel by pixel and divide them into three levels: completely inconsistent, partially consistent, and completely consistent. Only "completely consistent" areas are retained, which are considered to be areas with highly reliable classification results.

[0022] S3, using the geometric centers of the patches in the reserved area to construct sample points and mix them with the optimized samples to achieve sample quantity doubling and layout optimization; Specifically, the patch shape index (SI) was used to select patches of simple shape and moderate area from high-confidence areas of sample collection. The geometric centers of these patches were extracted and then merged with the high-quality samples obtained to double the sample size.

[0023] S4. Based on the superposition analysis of all unsupervised classification results and optimized samples, different sub-categories of samples in the same category are identified to complete the refined processing of sample types; S5. Based on the comparison of visual interpretation classification results, changes in class separation between samples, and changes in comparative accuracy evaluation, a comprehensive test is conducted to verify whether sample purification and optimization processing are effectively completed.

[0024] In this optional embodiment, an initial sample is selected based on visual interpretation of land cover types, an odd-numbered classifier is selected in the remote sensing image processing platform for supervised classification, and different land feature types are extracted, including: resampling and category mapping based on different types of land cover types, unifying the spatial resolution and land feature classification system; raster overlaying the intersection of different types of land cover types to construct a high-confidence land cover type initial sample; and an odd-numbered classifier is selected in the remote sensing image processing platform to perform supervised classification and extract classification results of several land feature types.

[0025] Specifically, unsupervised classification is performed on the unclassified remote sensing images in ENVI software, resulting in unsupervised classification results with 3n-5n categories (n is the number of training sample categories). For example, if the original training samples contain n main land feature types, the unsupervised classification results will further subdivide these land feature types into 3n to 5n subcategories. The optimized samples are superimposed on the unsupervised classification results. By comparing the sample category information with the category information in the unsupervised classification results, samples of the same category but in different classification categories are marked and divided into different subcategory samples. These subcategory samples are organized and annotated, and the sample dataset is updated to include more detailed sample type information, enabling refined processing of sample types.

[0026] In this optional embodiment, a classifier integration method is used to perform raster overlay on areas of different land feature types, and areas with completely consistent classification results are retained as retained areas; the initial training samples are superimposed on the retained areas, and samples outside the retained areas are screened and eliminated, completing the sample quality optimization processing, including: importing several classification results into the user desktop component, using a raster calculator to perform pixel-by-pixel overlay analysis on similar land features, and counting pixel classification consistency; based on the pixel classification consistency results, they are divided into completely inconsistent, partially consistent, and completely consistent, and all consistent areas are extracted as retained areas; the initial training samples are spatially overlaid with the retained areas to determine whether the samples are located within the high-confidence retained area; if the sample is within the retained area and is the same type as the regional land feature, the sample is retained, otherwise, the samples outside the retained area are screened out, and only consistent samples within the area are retained to optimize sample quality.

[0027] Specifically, select an odd number of classifiers (such as the Maximum Likelihood Classifier, Mahalanobis Distance Classifier, and Support Vector Machine Classifier) ​​in ENVI to perform supervised classification on the remote sensing image and obtain multiple classification results. Import the multiple classification results into ArcMap (the user desktop component) and use the Raster Calculator to perform pixel-by-pixel overlay analysis on these classification results. The specific steps are as follows: Consistency statistics: The consistency of the feature types in all classification results is counted pixel by pixel, and divided into three levels, including: Total inconsistency: The feature types in all classification results on the pixels are completely different. Partial consistency: The feature types in the pixels of the two classification results are the same. Total consistency: The feature types in the pixels of all classification results are exactly the same. Only "total consistency" areas are retained, which are regarded as high-confidence sample areas with highly reliable classification results. The existing training samples are superimposed and analyzed with the high-confidence sample areas. Through the spatial query function, the training samples outside the high-confidence area are filtered out and the training samples in the high-confidence area with consistent classification results are retained to complete the optimization processing of lower-quality samples.

[0028] In this optional embodiment, the geometric centers of the patches in the reserved area are used to construct sample points and mix them with the optimized samples to achieve sample quantity doubling and layout optimization, including: generating a grid of a preset format (divided into blocks according to the rule of 1000m×1000m) in the algorithm test area to achieve spatial blocking, and counting the number of samples in each grid to identify dense and sparse sample areas; extracting the feature vectors of each sample in the dense area, and calculating the similarity between samples based on the Euclidean distance, removing highly repeated samples, and retaining representative sample points; clustering the dense area samples using a spatial clustering algorithm, eliminating outliers and abnormal samples, and retaining a core sample point set; extracting the geometric centers of the high-confidence reserved area patches, generating sample points and merging them with the optimized dense area samples to achieve sample quantity doubling; using an unbalanced learning mechanism in the sparse area to complete the expansion of the number of samples in the sparse area.

[0029] Specifically, a 1000×1000m grid was generated on the image map to divide the test area, ensuring that the size of each grid was consistent. This process can be achieved in ArcMap by using the Create Fishnet tool. The number of samples in each grid was counted to determine areas with higher sample density (dense areas) and areas with lower sample density (sparse areas). For areas with higher sample density, sample quality screening and optimization were performed. This includes: calculating the feature vector of each sample point (such as spectral features, texture features, etc.). Calculating the similarity between sample points using methods such as Euclidean distance or cosine similarity. Sample points with too high similarity were removed, and the most representative sample points were retained. The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm was used to cluster the sample points, remove noise points and outliers, and retain the core sample points.

[0030] Specifically, the sample screening rules are dynamically adjusted in combination with the optimization algorithm, and the specific implementation is as follows: Genetic algorithms are used to optimize the sample screening rules, and the thresholds and conditions for sample screening are dynamically adjusted according to multi-dimensional information such as the spectral characteristics and spatial distribution of the samples. The weight parameters in the sample screening rules are adaptively adjusted according to the characteristics of different images, different scales and different regions to ensure the flexibility and adaptability of the screening rules. A sample quality assessment module is introduced to perform quality assessment on the screened samples to ensure the reliability and representativeness of the samples: by calculating the distance between the spectral feature vectors of the samples, the similarity of the spectral features of the samples is evaluated, and samples with too high similarity are removed. The spatial distribution of the samples is evaluated to ensure the uniform distribution of the samples in the algorithm test area to avoid concentrated or sparse distribution of samples.

[0031] Introducing a deep learning model to further screen and optimize samples: Using CNN to extract and classify sample features, combined with the high-precision classification capabilities of the deep learning model, further improves sample quality. Utilizing transfer learning technology, the features of existing high-quality samples are transferred to the current algorithm testing area, improving sample reliability and classification accuracy.

[0032] It needs to be explained that the specific implementation steps of using convolutional neural networks (CNN) for feature extraction and classification and utilizing transfer learning technology are explained in detail.

[0033] 1. Specific steps of CNN feature extraction and classification: 1) CNN is used as a dynamic quality assessment module, replacing the static thresholding rules of traditional overlay analysis to achieve adaptive sample selection. Image patches (e.g., 30×30 pixels) representing "all consistent regions" generated by multiple classifiers are used as positive samples (high-quality regions), while image patches representing "inconsistent regions" are used as negative samples (low-quality regions). Spectral features are combined with spatial context information, and the input data includes: The classification results (in one-hot encoding) of the multiple classifiers serve as spatial consistency labels.

[0034] 2) If Figure 2 As shown in the figure, the two-channel CNN network design: The spectral channel uses a 1D convolution kernel to process the band sequence and extract the spectral response pattern. The spatial channel uses a 3×3 convolution kernel to extract local texture and shape features. The output is a sample quality score (0-1), with samples >0.8 considered high quality.

[0035] 3) Dynamic screening rules: A CNN is trained using the initial sample library to generate a quality scoring model. Image blocks are generated for candidate samples and input into the CNN to obtain quality scores. Adaptive threshold adjustment: The threshold is dynamically adjusted based on image resolution (e.g., 10m / 30m) (the threshold for high-resolution images is increased to 0.85). This solves the problem of traditional overlay analysis misjudging complex feature boundaries (e.g., mixed pixels of bare soil and buildings).

[0036] 2. Specific steps of transfer learning technology: 1) Pre-trained model selection: Use ResNet50 pre-trained on ImageNet as the base model, retaining the convolutional layer weights.

[0037] 2) Two-stage transfer learning: Phase 1: Domain Adaptation: Sample image patches from multi-source LUCC data (GlobeLand30, FROM-GLC).

[0038] Phase 2: Region-Specific Optimization: High-confidence parent samples (initial samples) from the current algorithm test area are used. The bottom convolutional layers are frozen, and only the top fully connected layers are fine-tuned to avoid overfitting.

[0039] 3) Collaboration with Small Sample Optimization: For areas with few samples, we use the migrated CNN to generate synthetic samples. Feature Space Interpolation: We linearly interpolate the feature vectors of high-quality samples to generate new samples. Generative Adversarial Networks (GANs): We use a pre-trained generator to synthesize virtual samples with consistent spectral and spatial features.

[0040] Specifically, the algorithm test area is divided into several sub-areas of equal size according to the rule of 1000m×1000m, and the number of samples in each sub-area is counted. For areas with dense samples, further quality screening and optimization are performed. By calculating the distance between the characteristic vectors of the samples, samples with too high similarity are removed to improve the diversity of the samples. For sub-areas with sparse samples, the number of samples is multiplied by interpolation or generation of virtual samples according to the characteristics and distribution rules of the area to ensure that the sample distribution in each sub-area is more balanced. The optimized samples are overlaid with the original samples to ensure that the spatial distribution of the optimized samples is more uniform and reasonable. The specific operations include: counting the number of samples in each sub-area to ensure that the distribution of the optimized samples meets expectations. Through overlay analysis, the consistency of the spatial distribution of the optimized samples and the original samples is checked to ensure that the optimization process does not introduce new deviations. In order to further improve the accuracy and adaptability of overlay analysis, new steps and technical features are introduced in the overlay analysis process, as follows: In the overlay analysis process, adaptive weighting coefficient optimization is added. The specific operations are as follows: 1) Weight calculation: The weight of each classifier is dynamically calculated based on its classification accuracy. The higher the classification accuracy, the greater the weight. The weight calculation formula is as follows: ; in, w i It is i The weight of the classifier, P i It is i The classification accuracy of the classifier, n is the number of classifiers.

[0041] 2) Weighted Overlay Analysis: When performing overlay analysis, not only the consistency of feature types is considered, but also the weight of each classifier. The specific operation is as follows: The consistency of feature types across all classification results is statistically analyzed pixel by pixel, and a weighted sum is calculated based on the weights. Based on the weighted consistency statistics, high-confidence sample areas are extracted. Areas with higher weighted consistency statistics are considered more reliable.

[0042] In the overlay analysis process, add adaptive threshold adjustment, the specific operation is as follows: 1) Threshold calculation: The threshold is dynamically adjusted based on the sample distribution density and the confidence level of the classification results in each sub-region. The higher the sample distribution density, the lower the threshold; the higher the confidence level of the classification results, the higher the threshold. The threshold calculation formula is as follows: ; in, T i It isi The threshold of the sub-region, D i It is i The sample distribution density of the sub-regions is C i It is i The confidence of the classification results of the sub-regions, f is an adaptive function.

[0043] 2) Threshold Application: During overlay analysis, the threshold for each subregion is dynamically adjusted using an adaptive threshold adjustment mechanism. This includes extracting high-confidence sample regions based on the adaptive threshold. Only regions with weighted consistency statistics exceeding the threshold are considered high-confidence sample regions. By introducing adaptive weighting coefficient optimization and an adaptive threshold adjustment mechanism, parameters can be dynamically adjusted based on actual conditions during the overlay analysis process, improving analysis accuracy and adaptability.

[0044] In this optional embodiment, the geometric centers of the patches in the high-confidence retention area are extracted, sample points are generated and merged with the optimized dense area samples to achieve doubling of the sample quantity, including: using several classifiers to perform supervised classification on the remote sensing image, using the raster calculator in the user desktop component to superimpose multiple classification results, and extracting the high-confidence retention area with consistent classification results; in the high-confidence retention area, the shape index of each patch is calculated, and the patch shape indexes are arranged in ascending order, a preset amount of patches are extracted, and the target patches are retained; the geometric center tool of the user desktop component is used to extract the geometric centers of the screened patches, and the geometric center points of each patch are obtained as new sample points, and merged with the optimized dense area samples to achieve doubling of the sample quantity; quality assessment is performed on the new samples, and the sample separation and accuracy are calculated to ensure that the separation between the new samples and the original samples reaches a preset value.

[0045] Specifically, areas with identical feature types retained after raster overlay analysis were used as retained areas for sample expansion. The pixel areas in the retained areas were processed into patches of varying sizes and shapes, and smaller patches were removed by setting a threshold. The patch shape index (SI) was calculated for the area-screened feature patches and sorted in ascending order to extract patches with simple and concentrated shapes. The geometric center of each patch was obtained in ArcMap as a point sample. Subsequently, 50% of the point sample set for each type of feature was randomly selected as training samples for sample expansion, and 50% was used as evaluation samples. The effectiveness of sample expansion was evaluated by comparing the changes in sample size before and after optimization.

[0046] In this optional embodiment, the quality of the newly added samples is evaluated, and the sample separation and accuracy are calculated to ensure that the separation between the newly added samples and the original samples reaches a preset value, including: collecting spectral characteristics and category information of the newly added samples and the original samples to ensure that the data format is consistent; using the statistical distance formula to calculate the separation between samples of each category based on the sample feature mean and covariance; constructing a confusion matrix based on the supervised classification results of the newly added samples and the original samples, and using the confusion matrix to calculate the accuracy index, comparing the accuracy index before and after optimization, and evaluating the optimization effect.

[0047] Specifically, samples before and after optimization are used for classification to obtain classification results. A confusion matrix is ​​constructed based on the classification results. This includes: creating an n×n matrix, where n is the number of categories; counting the number of true and predicted samples for each category to populate the confusion matrix. Each sample is assigned an initial weight, determined based on its spectral characteristics and spatial location. During the multi-classifier ensemble process, sample weights are dynamically adjusted based on the classification results of each classifier. This includes: increasing the weights of samples consistently classified by multiple classifiers; and decreasing the weights of samples with significant disagreement between the classifiers. High-quality samples are screened based on the adjusted weights. This improves sample quality while maintaining sample quantity. The optimized samples are preprocessed to extract their spectral and spatial features. A convolutional neural network (CNN) model is used to train the samples and generate new sample features. Based on the generated features, new samples are generated using a generative adversarial network (GAN). This significantly increases the sample quantity while maintaining sample authenticity.

[0048] Specifically, use the same classification model and evaluation metrics to evaluate the accuracy of the classification results before and after sample optimization. Alternatively, use the training sample separation metric to analyze sample quality before and after optimization. This verifies that the sample purification and optimization have achieved the desired results. Compare the accuracy evaluation results before and after optimization. If there is a significant improvement in accuracy, the sample purification was successful. If there is no significant change or a decrease in accuracy, review each step of the optimization algorithm to identify any issues and make adjustments.

[0049] In this optional embodiment, the separation degree between samples of each category is calculated using a statistical distance formula based on the sample feature mean and covariance, including: calculating the mean and variance of each category sample matrix; calculating the sample matrices of different categories through Bhattacharyya distance, and calculating the separation degree between samples of each category using statistical distance.

[0050] Specifically, sample data before and after optimization was collected to ensure that the spectral characteristics and spatial location information of each sample were complete. The Jeffries-Matusita distance (statistical distance) formula was used to calculate the separation between samples of different categories. This included: for each category, the mean and variance of its sample matrix were calculated. The sample matrices of different categories were substituted into the Bhattacharyya distance formula to calculate the Bhattacharyya distance. Finally, the Jeffries-Matusita distance was calculated using the statistical distance formula. The calculated separation value was compared with a threshold of 1.8. A separation value close to 2.0 indicated good separability between samples and a significant optimization effect.

[0051] In this optional embodiment, an unbalanced learning mechanism is used in the sparse area to expand the number of samples in the sparse area, including: calculating the nearest neighbors between sample points in the sparse area; randomly inserting new sample points between the nearest neighbor sample points to generate synthetic samples, thereby expanding the number of samples in the sparse area.

[0052] Specifically, sample density statistics are used to identify areas with low sample density (sparse areas). For areas with sparse samples, a sample supplementation strategy based on an imbalanced learning mechanism is employed. This strategy uses the SMOTE (Synthetic Minority Over-sampling Technique) algorithm to generate synthetic samples. The SMOTE algorithm increases the number of samples in sparse areas by introducing synthetic samples into the feature space of minority class samples. This algorithm involves calculating the nearest neighbors between sample points in sparse areas; randomly inserting new sample points between the nearest neighbors to generate synthetic samples; and verifying the generated synthetic samples to ensure they conform to the actual ground features and avoid introducing excessive noise. During the sample supplementation process, an adaptive sample weight adjustment mechanism is introduced. The weights of sample points are dynamically adjusted based on their feature importance and distribution to ensure the quality and representativeness of the supplemented samples. Multi-scale feature fusion technology is also used to comprehensively analyze sample features at different scales. Multi-scale feature fusion can more comprehensively capture sample feature information, improving sample robustness and classification accuracy.

[0053] Specifically, first, the specific implementation steps of the adaptive sample weight adjustment mechanism include: 1) Weight initialization: In the dense area sample quality screening optimization stage, each sample point will be assigned an initial weight based on its spectral characteristics, spatial distribution characteristics, and classification confidence; the information entropy of each sample point is calculated as follows: ; in, E i For the i The information entropy of sample points, p kFor this sample point k The probability distribution of feature dimensions.

[0054] 2) Dynamically adjust weights based on information entropy: ; Here, λ is a smoothing factor (the default value is 0.1), which is used to avoid excessive weight when the information entropy approaches zero.

[0055] 3) In the sparse area sample supplementation stage, the weights are dynamically adjusted according to the classification contribution of the samples: Use a classifier (such as a support vector machine) to classify the samples and record the contribution of each sample to the classification accuracy. C i .

[0056] Update weights based on contribution: ; in, α is the weight gain coefficient (the default value is 0.5), the contribution C i It is calculated by the user accuracy (UA) in the confusion matrix.

[0057] 4) Specific application of weights in sample supplementation: When the SMOTE algorithm generates synthetic samples, samples with higher weights are preferred as parent samples; for samples with weights lower than the threshold (such as w i Samples with a value <0.3) are considered low-quality and are discarded. After each round of sample optimization, the sample weights are recalculated and fed back into the next round of optimization. Cross-validation is used to verify the effectiveness of weight adjustments and ensure that the weight mechanism is positively correlated with improved classification accuracy.

[0058] 2. Specific implementation steps of multi-scale feature fusion technology: 1) In the sample quality screening of dense areas, features are extracted from the following scales: Scale 1 (macroscale): Extract the spatial distribution density characteristics of samples based on a 1000m×1000m grid.

[0059] Scale 2 (mesoscale): Extract the spectral mean and texture features (such as gray-level co-occurrence matrix) of the samples based on a 500m×500m grid.

[0060] Scale 3 (microscale): Extract the spectral band reflectance and normalized difference vegetation index (NDVI) of the sample based on the pixel level (30m resolution).

[0061] 2) Fusion rules of multi-scale features: Construct a multi-scale feature vector, the formula is as follows: ; in, F macro 、 F meso 、 F micro Eigenvectors corresponding to the macro, meso and micro scales, respectively.

[0062] Perform weighted concatenation of features at different scales, with weights dynamically adjusted based on classifier feedback: ; in, β 1=0.3, β 2=0.5, β 3 = 0.2 (default value), which can be optimized based on training data.

[0063] 3) Application of multi-scale features in sample screening: In DBSCAN clustering, multi-scale fusion features are used instead of single spectral features to improve clustering accuracy. For synthetic samples generated in sparse areas, the similarity between their multi-scale features and real samples is verified, and synthetic samples with excessive differences are eliminated.

[0064] In this optional embodiment, based on the overlay analysis of all unsupervised classification results and optimized samples, and identifying different subcategory samples in the same category, the sample type refinement processing is completed, including: using the remote sensing image processing platform to perform unsupervised classification in the algorithm test area, and subdividing the ground object type into several multiple subcategories of the training sample category according to spectral characteristics; overlaying the optimized samples with the unsupervised classification results for analysis, and confirming the spatial distribution relationship of samples of the same category in different subcategories; marking the samples of the same category found after the overlay that are in different subcategory areas, and recording the corresponding spatial and category attributes; organizing and updating the sample attributes according to the marking results, completing the subcategory information assignment, and forming a refined sample data set containing several subcategories; using the hierarchical clustering algorithm to perform multi-level division of the subcategory samples to generate subcategories of the smallest granularity, and completing the sample type refinement processing.

[0065] Specifically, after quality screening, quantity expansion, and layout optimization of the samples, ENVI was used to perform unsupervised classification on the algorithm test area. Ground objects in the test area were initially divided based on the spectral information within the image. The number of unsupervised classification categories was set to at least five times the number of training sample categories, based on the number of original sample types (for example, if the original training sample contained n major ground object types, the unsupervised classification results would further subdivide these ground object types into at least 5n subcategories). The optimized samples were overlaid with the unsupervised classification results. By comparing the sample category information with the category information in the unsupervised classification results, samples with the same category but different classification results were labeled and divided into different subcategories. This process involved matching the sample categories after supervised classification with the categories in the unsupervised classification results, and identifying the different subcategories of samples of the same category in the unsupervised classification. These subcategory samples were labeled, and their location and category information were recorded. These subcategory samples were organized and annotated, and the sample dataset was updated to include more detailed sample type information.

[0066] To further refine the classification and identification of subcategory samples, a hierarchical clustering algorithm was introduced to perform a multi-level decomposition of the unsupervised classification results. This involves using a hierarchical clustering algorithm to cluster the unsupervised classification results, forming a multi-level subcategory classification. Hierarchical clustering can reveal structural characteristics at different levels, facilitating more detailed identification of feature types. The clustering results are then decomposed at multiple levels, further subdividing each subcategory into smaller subcategories. This allows for more precise capture of subtle differences between feature types. Based on the hierarchical clustering results, subcategory samples are merged and updated to ensure a more refined and accurate classification of each subcategory.

[0067] Based on unsupervised classification, a hierarchical clustering algorithm is introduced to optimize subcategories. Hierarchical clustering not only reveals structural features at different levels but also helps identify feature types more precisely. Multi-level decomposition allows each subcategory to be further subdivided into smaller subcategories, resulting in a more refined sample partitioning. Multi-level decomposition is performed on the clustering results, further subdividing each subcategory into smaller subcategories. This approach more accurately captures subtle differences between feature types and improves classification accuracy. Based on the hierarchical clustering results, subcategory samples are merged and updated to ensure a more refined and accurate partitioning of each subcategory. This improves sample representativeness and further enhances the accuracy and stability of the classification model. Multi-level decomposition of subcategories allows for more precise identification of subtle differences between feature types, thereby improving classification accuracy. Merging and updating subcategory samples enhances sample representativeness and ensures that the classification model better learns the distribution patterns of the data. Multi-level decomposition adapts to feature types of varying complexity, resulting in greater adaptability and robustness.

[0068] It is necessary to explain the logical relationship of the four-step optimization operation of the present invention: 1) Sample quality screening: Remove low-quality training samples to ensure that the samples processed subsequently have high credibility. Initial training samples (including samples interpreted by manual visual interpretation). Select an odd number of classifiers in ENVI to perform supervised classification of remote sensing images to obtain multiple classification results. Use the raster calculator in ArcMap to extract different land feature types from all classification results. Use the idea of ​​multi-classifier integration to perform raster overlay analysis on areas with the same land feature type, and retain areas with completely consistent classification results. Superimpose the initial training samples with the reserved area, remove samples outside the reserved area, and retain samples within the reserved area that are exactly the same as the regional land feature type. A high-quality training sample set reduces the interference of low-quality samples.

[0069] 2) Sample Expansion: While maintaining sample quality, the sample size is increased to improve the classifier's generalization and classification accuracy. High-quality training samples are screened for quality. Geometric centers are extracted from patches in the retained area, and these centers are used as point samples. A portion of the point sample set for each type of feature is randomly selected as additional training samples, and another portion is used as evaluation samples. This ensures the validity and representativeness of the additional samples. The expanded high-quality training sample set significantly increases in number.

[0070] 3) Sample Layout Optimization: Ensure a more even spatial distribution of samples within the algorithm testing area, avoiding overly concentrated or sparse samples and improving the stability and accuracy of classification results. High-quality training samples after quantity expansion. Divide the algorithm testing area into several equally sized sub-areas. Perform quality screening in densely sampled areas, removing samples with excessive similarity to increase sample diversity. In sparsely sampled areas, expand the sample quantity to ensure a more balanced distribution within each sub-area. The optimized layout results in a more evenly distributed set of high-quality training samples.

[0071] 4) Sample Type Refinement: Further refine sample categories to provide more accurate feature information for the classifier and improve classification accuracy. Optimize the layout of high-quality training samples. Perform unsupervised classification on the algorithm test area in ENVI, dividing the ground type in the algorithm test area into multiple subcategories. Overlay the optimized samples with the unsupervised classification results, labeling and classifying samples within the same category that fall into different subcategories. Update the sample dataset to include more detailed sample type information. This refined set of high-quality training samples features even more detailed sample categories.

[0072] The technical route of the present invention is as follows Figure 3 As shown, what needs to be explained is: 1. Research content and ideas: First, appropriate training samples were selected based on the different land cover types in different urban areas through manual visual interpretation. In ENVI (a remote sensing image processing platform), an odd number of classifiers were selected to simultaneously perform supervised classification on multiple remote sensing images. In ArcMap (a user desktop component), the raster calculator was used to extract the different land feature types from all classification results. Using the concept of multi-classifier integration, raster overlay analysis was performed on areas with identical land feature types. Areas with identical classification results were retained to serve as reference areas for sample quality screening and as candidates for sample expansion. Initial training samples marked by manual visual interpretation were superimposed with the retained area. Initial training samples outside the retained area were screened and removed. Initial training samples within the retained area that were identical to the regional feature type were retained, completing the screening and removal of low-quality samples. The geometric centers of the patches in the retained area were selected and retained as point data samples, which were then mixed with the optimized samples to achieve an optimization process that doubled the number of samples. The algorithm test area was divided into blocks, with quality screening and optimization performed in areas with dense samples and sample expansion in areas with sparse samples, completing the sample layout optimization process. Unsupervised classification was performed on all urban areas, and the optimized samples were superimposed with the unsupervised classification results. Samples with different unsupervised classification results within the same supervised classification category were considered to be different subcategories of the same category, completing the refinement of sample types. Finally, the successful purification of samples was verified by comparing the visual interpretation classification results, changes in class separation between samples, and changes in comparative accuracy evaluation.

[0073] II. Algorithm Test Data and Processing Methods: The algorithm test area is a province located between 31°23'-36°22'N and 110°21'-116°39'E, with a total area of ​​approximately 167,000 square kilometers. The province's terrain rises gradually from north to south and descends gradually from east to west, consisting of plains and basins, mountains, hills, and water surfaces. It spans the four major river basins of the Haihe, Yellow, Huaihe, and Yangtze Rivers, with an average total water resource of approximately 40.5 billion cubic meters. The algorithm test area data was selected from Landsat 8 satellite digital remote sensing imagery (7 bands) of various urban areas in the province, with an image size of 4403 pixels by 4474 pixels. The imaging date was March 19, 2020. Landform types within the algorithm test area include water bodies (lakes, rivers), built-up land (urban and rural areas), bare soil, forest land, cultivated land (including various crops), and mountains. After correcting the image using radiometric calibration and atmospheric correction, the data of each urban area in the algorithm test area are mosaicked and the boundaries are segmented using their respective urban boundary lines to complete the data preprocessing.

[0074] 3) Sample data processing related methods: Overlay analysis is a spatial analysis method widely used in GIS. Based on the same coordinate system, datum, and area, different data layers involving the subject are overlaid together to generate a new data layer. The generated layer integrates all the attributes of the original data layer and establishes new attribute relationships. From a computational point of view, overlay analysis is a method of performing spatial logical operations on two or more data layers, including ENVI intersection, union, and difference operations. For raster data, ENVI will generate a new raster data, in which the value of each pixel is obtained by calculating the original raster data involved in the calculation. The patch shape index SI (Shape Index) is a calculation method that measures the complexity of a shape by calculating the degree of deviation between a certain shape and a circle or square of the same area. There are two forms of common patch index S: Plaque circularity index: ; Patch Squareness Index: ; When the index value is low, the patch appearance will be more regular, simple and closer to the geometric shape of a square or circle; as the value continues to increase, the patch appearance will become more complex, and the difference between its length and width will become more obvious. The mean center refers to the average x-coordinate and y-coordinate of all geometric elements in the algorithm test area, which is usually used to track distribution changes and compare the distribution of different types of elements. By calculating the average coordinates of geometric elements, the characteristics and trends of different types of spatial distributions are displayed. The mean center can be expressed as: ; in x i and y i It is an element i The coordinates of n is the sum of the elements.

[0075] 4) Sample Optimization Result Evaluation Method: Sample Separation: The degree of difference between sample categories determines whether the classifier can accurately classify the pixels in the image into the specified feature category. A metric that measures the separability or distance between samples is needed to determine the differences between different types of samples. The Jeffries-Matusita distance is currently the primary metric for evaluating sample separation. For sample data that follows a normal distribution, the JM distance (statistical distance) formula is as follows: ; ; Where, J ij Represents category samplesi With category samples j The degree of separation between B ij Represents category samples i With category samples j Bhattacharyya distance between m i and m j Represents the mean of the sample matrix of two categories; C i and C j Represents the variance of the sample matrix of two categories; T represents the transposed matrix; represents the inverse matrix; represents its determinant; e Represents a natural constant. The result ranges from 0 to 2.0. A distance closer to 2 indicates that the accuracy of classifying any pixel into one of the two categories is closer to 100%. Confusion Matrix: The most commonly used method for remote sensing image classification accuracy assessment is the confusion matrix. The confusion matrix, also known as the error matrix, is a standard format that uses an n-row, n-column matrix to represent accuracy assessment results. The confusion matrix provides a clear understanding of the classification results for each category and allows for accuracy assessment.

[0076] ; In the confusion matrix formula, each column of the confusion matrix represents the predicted category, and the total number of each column represents the number of data predicted to be of that category; each row represents the true category of the data, and the total number of data in each row represents the number of data instances of that category. P i，j Indicates that it will belong to the true category i Classify the image into the predicted category j The total number of images, n is the number of categories. P i，j As the proportion of gradually increases, the accuracy of image classification will continue to improve. The confusion matrix includes producer accuracy, customer satisfaction, overall accuracy and Kappa coefficient (reliability) which are key factors. Together they constitute the four benchmarks for measuring accuracy. Producer accuracy: Producer accuracy (PA), also known as cartographic accuracy, refers to the accuracy of the classifier correctly classifying the sample into the first j The number of pixels of the class j The total number of true references for each class (the jThis data can be used to evaluate whether the classification results of a specific location are consistent when a random sample is extracted from the ground dataset even without reliable evidence. In the confusion matrix, it is calculated as follows: ; Where, n is the number of categories; P i，j Indicates that it will belong to the true category i Classify the image into the predicted category j User accuracy: User accuracy (UA) can be measured by calculating the ratio of the number of correctly classified pixels of a certain category to the total number of pixels of that category. The confusion matrix method is used to reflect the accuracy of the sample, which indicates the degree of consistency between the samples extracted from the classification results and the actual types on the ground. The calculation formula for user accuracy is: ; Overall accuracy: Overall accuracy (OA) is a statistic with probabilistic significance, which refers to the percentage of all correctly classified land cover classification points to the total number of sampling points; that is, the sum of all values ​​on the diagonal of the confusion matrix divided by the sum of all samples. Its calculation formula in the confusion matrix is: ; Kappa Coefficient: The Kappa metric is a measure of how well a classification result compares to random chance. The Kappa coefficient can be estimated using the following equation. It can be calculated using the following formula: ; in, r Indicates the total number of categories; x ii It refers to the number of correctly classified samples in the confusion matrix; x i+ Indicates the i The total number of true samples of the class, x +i It refers to the i The total number of predicted samples of the class; N It refers to the overall evaluation of the model, that is, the model contains the total amount of each element of the model. It can fully consider the characteristics of the chaotic matrix and apply it to measure the accuracy of classification.

[0077] 3. Analysis of sample optimization results: 1) Sample question analysis: 1.1) Sample quality: When training samples are of low quality, there are several reasons: Problem of imbalance in the number of different samples: When the number of samples of a certain type of land object is significantly less than that of other types, the problem of imbalance in the number of categories will occur. During the classification process, the classifier will be more inclined to land object types with sufficient number and sufficient features. When the number of samples is small and cannot provide more features, the classification effect of this type of land object will be reduced, resulting in a decrease in the overall classification accuracy; Data collection problem: If the training samples are obtained through field collection, errors may occur in the data collection process, resulting in low quality of training samples; Human problem: When the initial training samples are obtained by manual visual interpretation, the quality of training samples will vary due to the interpreter's familiarity with the algorithm test area and different interpretation experience; Spectral line feature similarity problem: Different land objects have similar spectral curves due to the influence of growth environment, pests and diseases, and radioactive substances, resulting in poor quality of training samples, which easily leads to misclassification and misclassification in the separation results. Therefore, the present invention proposes a "sample quality screening" method to solve the problem of low sample quality. In ENVI (a remote sensing image processing platform), supervised classification of remote sensing images is performed using an odd number of classifiers. The classification results are then used in ArcMap (a user desktop component) to extract different feature types using a raster calculator. Using the concept of multi-classifier integration, raster overlay analysis is performed on the classification results of the same feature, retaining areas with exactly the same feature types. Each type of training sample and the reserved area are overlaid with the corresponding category. If the sample is within the reserved area and has the same feature type as the regional area, the sample is retained; if the sample is within the reserved area but has a different feature type than the reserved area, the sample is filtered out; if the sample is outside the reserved area, the sample is also filtered out, completing the screening and optimization of sample quality.

[0078] 1.2) Sample Quantity: When the number of training samples is limited, there are several reasons: Limited data availability: The process of obtaining a large and diverse number of training samples may be limited by factors such as small labeled areas and challenging collection processes, which may limit the number of available samples; The scarcity of certain types of features: When acquiring samples, certain types of features in the algorithm test area may be scattered and less numerous than other types, making it difficult to label them in large quantities, resulting in a limited number of samples; Cost and workload: Field collection or manual sample labeling can be labor-intensive, time-consuming, and labor-intensive, and these constraints may result in a limited number of samples. Therefore, this paper proposes a "sample expansion" method to address the issue of limited sample quantities. Areas with identical feature types, as determined by raster overlay analysis, are retained for sample expansion. The pixel areas in these retained areas are then processed into patches of varying sizes and shapes. First, small patches are removed by setting a threshold to filter out the smaller patches. In order to more accurately reflect the types of samples and the spatial layout of each type of sample, this paper calculates the patch shape squareness index of the feature patches after area screening and arranges them in ascending order, extracts the top 30% of the patch index of each feature, screens out narrow or overly complex patches, obtains sample areas with concentrated pixel distribution and simple shapes, obtains the geographic center of each feature patch in ArcMap as a point sample to replace the feature patch, and then randomly selects 50% of the point sample set of each type of feature as training samples for quantity expansion, and 50% as evaluation samples for quantity expansion, completing the optimization of sample quantity expansion.

[0079] 1.3) Sample layout: If the existing samples are unevenly distributed in the algorithm test area, there are several reasons: Bias in the sampling process: When the training samples are sampled manually, they are not randomly or systematically sampled across the entire space, which may result in an excessive or insufficient number of training samples in some areas, leading to an uneven spatial distribution of samples; Restrictions on sample collection conditions: Field sample collection may be subject to factors such as the collection path, sensor coverage, and data acquisition permissions, which may result in different levels of difficulty in different regions, which may result in a large number of training samples in some areas and a small number of training samples in some areas, resulting in an uneven spatial distribution; Specific land use patterns: In certain specific cases, the special spatial characteristics of some land features may cause the land features to be concentrated or dispersed in a certain area, which may result in an uneven spatial distribution of the collected samples. Therefore, the present invention proposes a "thick screening and small increase" method to solve the problem of uneven spatial distribution of samples. A 1000×1000m grid was generated on the image to divide the test area. Based on the sample quality screening and sample quantity expansion methods, the sample quality screening optimization operation was performed for areas with dense sample numbers, and the samples with the best quality and the most representative were selected. For areas with sparse sample numbers, the sample quantity expansion optimization operation was performed while ensuring the original sample quality.

[0080] 1.4) Sample Variety: When the algorithm test area has a wide variety of land cover, the sample classification may be too coarse. This can be due to several reasons: Limited theoretical knowledge: If knowledge or understanding of the specific ground objects in the algorithm test area is limited, labeling and collecting training samples that accurately represent the various categories can be challenging. Insufficient understanding of the diversity of ground objects can lead to overly coarse classifications in the training samples. Simplified category definitions: Using overly broad or generalized category definitions can result in coarse training samples. Simplified category definitions may not adequately represent the differences between different types of ground objects within the same category. Insufficient field data: If field data collection is limited in terms of spatial coverage, sample size, and sample variety, the representation of ground objects in the training samples may be too coarse. Therefore, this paper proposes a "sample refinement" method to address this issue. After sample quality screening, sample expansion, and layout optimization, unsupervised classification is performed on the algorithm test area using ENVI. Ground objects in the test area are divided based on the spectral information in the imagery. The number of unsupervised classification categories is set based on the number of existing sample types. The optimized samples are superimposed on the results of the unsupervised classification. If different classification categories in the unsupervised classification fall into the same class, then these different class samples are different subclass samples of the same type of feature. After using the subclass samples for supervised classification, the subclass features of each class are merged to complete the refined processing of sample types.

[0081] It should be explained that, in order to verify the accuracy and reliability of the method of the present invention, the area covered by the LC81230402020360LGN00 remote sensing image is used as the research object to verify the proposed training sample optimization method.

[0082] The analysis is conducted from three aspects: sample quantity, sample layout and sample quality: Sample size: According to Table 1 and Figure 4 It can be seen that the number of training samples doubled after purification, and the total number of purified samples reached 25.1 times that of the initial samples, indicating that the purification method is effective.

[0083] Table 1. Statistics of land use samples in multiple sample sets Sample layout: Figure 5 ( Figure 5 The left side of the figure is the initial sample distribution diagram) and Figure 5 ( Figure 5 The distribution map of sample set 3 is on the right side of the middle panel. The comparison shows that the purified sample points are more evenly distributed in space. The units in the figure are miles.

[0084] Sample quality: As shown in Table 2, it is intuitively shown that the quality of training samples has been improved and the category boundaries are clearer. Figure 6 It was found that most of the training samples were consistent with the types of land features presented by remote sensing images. Figure 7 and Figure 8 It shows that the optimized sample set 3 has a better overall classification accuracy (such as Figure 7 ) and Kappa coefficient (as shown in Figure 8 (as shown in Figure 2) are all better than the initial samples, highlighting the role of sample optimization.

[0085] Table 2. Statistics of sample separation of training samples Select the CLCD land cover dataset of the study area (such as Figure 9 ) as a reference benchmark, compared Figure 9 and Figure 10 Verification and optimization of samples achieve more refined land feature classification, effectively reduce misclassification, and make land type distribution more consistent with actual geographical logic.

[0086] right Figure 11 and Figure 12 Visual interpretation revealed that typical misclassifications in the original classification scheme were effectively improved, and some classification deviations were corrected after refined processing. The above experiments fully demonstrated the effectiveness of the sample optimization method proposed in this invention.

[0087] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A remote sensing image classification training sample optimization method based on ensemble learning, characterized in that: include: Initial samples were selected based on visual interpretation of land cover types, and an odd number of classifiers were selected in the remote sensing image processing platform to supervise the classification and extract different land feature types; The classifier ensemble method is used to overlay the grids of regions with different land feature types, and the regions with completely consistent classification results are retained as the reserved regions. The initial training samples are superimposed on the reserved area, and the samples outside the reserved area are screened and eliminated to complete the sample quality optimization process; The geometric centers of the patches in the reserved area are used to construct sample points and mixed with the optimized samples to achieve sample quantity multiplication and layout optimization; Based on the superposition analysis of all unsupervised classification results and optimized samples, different sub-categories of samples in the same category are identified to complete the refined processing of sample types; Based on the comparison of visual interpretation classification results, changes in class separation between samples, and changes in comparative accuracy evaluation, a comprehensive test was conducted to verify whether sample purification and optimization processing were effectively completed.

2. The remote sensing image classification training sample optimization method based on ensemble learning according to claim 1, characterized in that: The initial samples are selected based on the visual interpretation of land cover types, and an odd number of classifiers are selected in the remote sensing image processing platform for supervised classification, and different land feature types are extracted, including: Resampling and category mapping based on different types of land cover to unify spatial resolution and land feature classification system; Perform grid overlay on the intersection of different land cover types to construct an initial sample of high-confidence land cover types; An odd number of classifiers are selected in the remote sensing image processing platform to perform supervised classification and extract classification results of several land feature types.

3. The remote sensing image classification training sample optimization method based on ensemble learning according to claim 1, characterized in that: The method of using classifier integration to perform grid overlay on regions of different land feature types, and retaining regions with completely consistent classification results as retained regions; The initial training samples are superimposed on the reserved area, and the samples outside the reserved area are screened and eliminated. The sample quality optimization process includes: After importing several classification results into the user desktop component, use the raster calculator to perform pixel-by-pixel overlay analysis on similar features and calculate the consistency of pixel classification; According to the pixel classification consistency results, the pixels are divided into completely inconsistent, partially consistent and completely consistent areas, and the completely consistent areas are extracted as the retained areas; Perform spatial overlay analysis on the initial training samples and the reserved area to determine whether the samples are located within the high-confidence reserved area; If the sample is located in the reserved area and has the same type as the regional feature, the sample is retained. Otherwise, the samples outside the reserved area are screened out and only the consistent samples in the area are retained to optimize the sample quality.

4. The remote sensing image classification training sample optimization method based on ensemble learning according to claim 1, characterized in that: The method of constructing sample points by using the geometric centers of the patches in the reserved area and mixing them with the optimized samples to achieve sample quantity multiplication and layout optimization includes: Generate a preset grid in the algorithm test area to achieve spatial partitioning, count the number of samples in each grid, and identify areas with dense and sparse samples; Extract the feature vectors of each sample in the dense area, calculate the similarity between samples based on Euclidean distance, remove highly repeated samples, and retain representative sample points; A spatial clustering algorithm is used to cluster samples in dense areas, eliminating outliers and abnormal samples and retaining the core sample point set; Extract the geometric center of the high-confidence retained area patches, generate sample points and merge them with the optimized dense area samples to double the number of samples; An unbalanced learning mechanism is used in sparse areas to expand the number of samples in the sparse areas.

5. The remote sensing image classification training sample optimization method based on ensemble learning according to claim 4 is characterized in that: The step of extracting the geometric center of the high confidence retained area patches, generating sample points, and merging the sample points with the optimized dense area samples to achieve doubling of the sample quantity includes: Use several classifiers to perform supervised classification on remote sensing images, use the raster calculator in the user desktop component to overlay multiple classification results, and extract high-confidence retention areas with consistent classification results; In the high-confidence retention area, the shape index of each plaque is calculated, and the plaque shape indexes are arranged in ascending order, a preset number of plaques are extracted, and the target plaque is retained; Use the geometric center tool of the user desktop component to extract the geometric center of the screened patches, obtain the geometric center point of each patch as a new sample point, and merge it with the optimized dense area sample to double the number of samples; Perform quality assessment on newly added samples, calculate sample separation and accuracy, and ensure that the separation between the newly added samples and the original samples reaches the preset value.

6. The method for optimizing remote sensing image classification training samples based on ensemble learning according to claim 5, characterized in that: The quality assessment of the newly added samples, calculation of the sample separation and accuracy, and ensuring that the separation between the newly added samples and the original samples reaches a preset value include: Collect the spectral characteristics and category information of the newly added samples and the original samples to ensure the consistency of the data format; According to the sample feature mean and covariance, the statistical distance formula is used to calculate the separation between samples of each category; According to the supervised classification results of the newly added samples and the original samples, a confusion matrix is ​​constructed, and the accuracy index is calculated using the confusion matrix. The accuracy index before and after optimization is compared to evaluate the optimization effect.

7. The method for optimizing remote sensing image classification training samples based on ensemble learning according to claim 6, characterized in that: The calculation of the separation between samples of each category using the statistical distance formula based on the sample feature mean and covariance includes: Calculate the mean and variance of the sample matrix for each category; The sample matrices of different categories are calculated by Bhattacharyya distance, and the separation degree between samples of different categories is calculated using statistical distance.

8. The method for optimizing remote sensing image classification training samples based on ensemble learning according to claim 7, characterized in that: The statistical distance calculation formula is: ； ； Where, J ij Represents category samples i With category samples j The degree of separation between B ij Represents category samples i With category samples j Bhattacharyya distance between them; m i and m j Represents the mean of the sample matrix of two categories; C i and C j Represents the variance of the sample matrix of two categories; T represents the transposed matrix; represents the inverse matrix; e Represents a natural constant.

9. The method for optimizing remote sensing image classification training samples based on ensemble learning according to claim 4, characterized in that: The use of the imbalanced learning mechanism in the sparse area to expand the number of samples in the sparse area includes: Calculate the nearest neighbors between sample points in the sparse area; New sample points are randomly inserted between the nearest neighbor sample points to generate synthetic samples and expand the number of samples in the sparse area.

10. The remote sensing image classification training sample optimization method based on ensemble learning according to claim 1, characterized in that: The above-mentioned method is based on the superposition analysis of all unsupervised classification results and optimized samples, and identifies different sub-category samples in the same category to complete the refined processing of sample types, including: In the algorithm testing area, the remote sensing image processing platform was used to perform unsupervised classification, and the ground object types were subdivided into several subcategories of the training sample categories based on spectral characteristics; The optimized samples are superimposed and analyzed with the unsupervised classification results to confirm the spatial distribution relationship of samples of the same category in different subcategories; Mark the samples of the same type found after superposition that are in different subcategory areas and record the corresponding spatial and category attributes; According to the labeling results, the sample attributes are sorted and updated, and the subcategory information is assigned to form a refined sample data set containing several subcategories; The hierarchical clustering algorithm is used to divide the subcategory samples into multiple levels, generate subcategories with the smallest granularity, and complete the refined processing of sample types.

Images