Remote sensing image secondary feature fine mapping method based on adaptive discriminative enhancement

Abstract

The application discloses a remote sensing image secondary feature fine mapping method based on adaptive discriminant enhancement, belongs to the field of remote sensing information processing and geographic information science, and comprises the following steps: selecting sample points of target secondary classification features and preprocessing the sample points; initializing an optimal feature set and a candidate feature set; judging whether a loop termination condition is met; if the loop termination condition is not met, then traversing each band in the candidate feature set, calculating ACSI values of the bands, and storing the ACSI values in a temporary candidate list; sorting the ACSI values in the temporary candidate list to obtain a band with the highest ACSI value, updating the optimal feature set and the candidate feature set by using the band; terminating the loop, outputting bands in the optimal feature set in a stacking order, and inputting the bands into a machine learning model to generate a secondary feature fine classification map. The application fundamentally solves the problem that the prior art cannot effectively retain weak discriminant signals in remote sensing fine classification.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of remote sensing information processing and geographic information science, and particularly relates to a remote sensing image secondary feature fine mapping method based on adaptive discriminant enhancement. BACKGROUND

[0002] In ecological monitoring, precision agriculture and urban planning applications, accurately distinguishing between secondary feature categories with highly similar spectral characteristics (such as distinguishing between coniferous forest and broadleaf forest, rather than primary features such as vegetation, water, and cities), is a long-standing technical bottleneck. When continuous time series images cannot be obtained due to weather conditions (such as clouds and rain), or expensive multi-source data (such as synthetic aperture radar and high spatial resolution optical data) cannot be used, how to efficiently extract the most discriminative information from single-phase multispectral images in data-sparse areas becomes the key to success and a crucial link to ensuring timely and periodic business monitoring.

[0003] Existing single-phase image feature selection methods mainly focus on balancing the "maximum class discrimination" and "minimum information redundancy" objectives. Among them, the most representative and technically similar advanced method is the Minimum Redundance Maximum Relevance (mRMR) algorithm. The mRMR algorithm is based on mutual information criteria and attempts to simultaneously optimize two objectives: one is to maximize the correlation between each feature (band) and the final classification label, and the other is to minimize the redundancy between features. In bioinformatics and remote sensing primary classification tasks, the mRMR algorithm has been widely used. However, when dealing with fine remote sensing secondary classification tasks, the mRMR and similar methods have exposed significant defects:

[0004] 1. Over-punishment of weak signals: The essence of fine classification is often to capture key signals composed of subtle differences in reflectance values on specific bands in the context of overall high correlation spectra. The "correlation-redundancy" balancing mechanism of mRMR imposes excessive punishment on bands with similar spectral trends (i.e., high redundancy), which easily leads to the incorrect filtering of bands that contain decisive discriminative information, resulting in permanent loss of effective information.

[0005] 2. Limitations of linear balance mode: The linear balance or simple ratio mode used by mRMR cannot adapt to the complex nonlinear relationships of remote sensing signals. It assumes that the importance of signals and the harmfulness of redundancy are of the same order of magnitude, but in fine classification, since they belong to the same large category, their spectral reflection curves must be similar in waveform, and the value of the unique discriminative signal that reflects their differences may far exceed the negative impact of the redundancy it brings.

[0006] 3. Mismatch in application scenarios: General feature selection algorithms such as mRMR were initially used in fields such as gene sequence analysis and text classification. Their design did not fully consider the physical characteristics of spectral signals in remote sensing images and the special challenges of fine classification tasks.

[0007] Therefore, there is an urgent need in this field for a novel feature selection method specifically designed for refined classification of remote sensing. It must be able to overcome the limitations of existing algorithms and intelligently identify and protect those key discriminative signals hidden in a strongly correlated background. Summary of the Invention

[0008] To address the aforementioned shortcomings in existing technologies, this invention provides a method for refining secondary ground features in remote sensing images based on adaptive discriminative enhancement. It fundamentally proposes a nonlinear, signal-priority discriminative information measurement paradigm that can intelligently amplify and retain the most critical feature signals in remote sensing classification tasks, thus overcoming the limitations of existing algorithms.

[0009] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for refining secondary ground cover mapping of remote sensing images based on adaptive discriminative enhancement, comprising the following steps:

[0010] S1: Select sample points of the target secondary classification land cover and preprocess them to obtain land cover category label samples;

[0011] S2: Based on the land cover category label samples, define the spectral band set and initialize the optimal feature set and candidate feature set;

[0012] S3: Based on the optimal feature set and the candidate feature set, determine whether the loop termination condition is met. If not, traverse each band in the candidate feature set, calculate the ACSI value of each band, and store it in the temporary candidate list. Then proceed to step S4. Otherwise, proceed to step S5.

[0013] S4: Sort the ACSI values ​​in the temporary candidate list, obtain the band with the highest ACSI value, update the optimal feature set and candidate feature set using the band, clear the temporary candidate list, and return to step S3.

[0014] S5: The loop terminates, and the bands in the optimal feature set are output in the order they were pushed onto the stack. Each band is then input into the machine learning model to generate a secondary land cover fine classification map.

[0015] Further, step S1 includes the following sub-steps:

[0016] S11: Select sample points according to preset rules;

[0017] S12: converting the sample point into an intersection point with the same resolution as the image grid according to the affine transformation matrix of the remote sensing image;

[0018] S13: expanding n grid points outward from the intersection point as the center;

[0019] S14: calculating the grid mean value for each spectral channel of the image in the expanded grid as the spectral feature curve of the sample point.

[0020] Further, the S2 comprises the following steps:

[0021] S21: inputting remote sensing image data and ground object category label samples, defining a spectral band set, and initializing an optimal feature set and a candidate feature set, wherein the optimal feature set is an empty set;

[0022] S22: traversing the tunable amplification operator of all candidate bands in the candidate feature set, and selecting the band with the highest tunable amplification operator value;

[0023] S23: removing the band with the highest tunable amplification operator value from the candidate feature set and adding it as an initial feature to the optimal feature set, to complete the algorithm initialization.

[0024] Further, the calculation of the tunable amplification operator comprises the following steps:

[0025] S221: for each candidate band in the candidate feature set, calculating the mean value of each ground object category in the candidate band and the mean value of the sample set in the candidate band;

[0026] S222: connecting the mean values of all sampling sample points of each ground object category in each band to form the spectral feature vector of the category;

[0027] S223: calculating the Pearson correlation coefficient of the spectral feature vectors of any two categories of ground objects, and calculating the average correlation coefficient of the two combinations according to the categories to obtain the overall correlation coefficient;

[0028] S224: determining the adjustable non-linear amplification index according to the overall correlation coefficient, and calculating the tunable amplification operator value based on the adjustable non-linear amplification index, the inter-class dispersion and the intra-class dispersion of the band.

[0029] Further, the adjustable non-linear amplification index calculation formula is:

[0030]

[0031]

[0032]

[0033] in, To provide an adjustable nonlinear amplification index, The overall correlation coefficient is... To determine whether it belongs to a segmentation point of fine-grained classification, The number of land cover categories for fine-grained classification. The Pearson correlation coefficient is used. For the first Spectral feature vectors of land-like features For the first Spectral feature vectors of land-like features For the first Land cover samples in characteristic bands The mean of the above, For the first Spectral mean of land-like features across the entire band. For the first Land cover samples in characteristic bands The mean of the above, For the first Spectral mean of land features across the entire band This represents the total number of bands.

[0034] Furthermore, in step S3, traversing each band in the candidate feature set and calculating the ACSI value for each band includes the following sub-steps:

[0035] S31: For each feature band in the candidate feature set Iterate through each selected band in the current optimal feature set. The sample values ​​of each band are concatenated and vectorized to obtain the spectral feature vector, and the feature bands are calculated. With each band Similarity;

[0036] S32: Determine the maximum similarity and calculate the characteristic bands. The information gain coefficient relative to the current optimal feature set;

[0037] S33: Calculate the characteristic band based on the information gain coefficient and the tunable amplification operator. ACSI value.

[0038] Furthermore, the similarity The calculation formula is:

[0039]

[0040] in, , The first Individual ground feature sample points in the band With band Sample values ​​on The total number of samples.

[0041] Furthermore, the characteristic band The ACSI value is:

[0042]

[0043] in, For band ACSI value, For the first A collection of land cover samples. For the first Number of samples corresponding to land cover types For the sample set in the characteristic band The mean of the above, for any sample In characteristic bands The value on, This is the set of feature bands that have been selected into the optimal feature set, and the set length is <= .

[0044] Furthermore, the loop termination condition includes:

[0045] (1) The length of the optimal feature set reaches the preset length;

[0046] (2) The band with the highest ACSI value is designated as the optimal candidate band, and the ACSI value of the optimal candidate band is lower than a preset contribution threshold.

[0047] The loop terminates when any of the above conditions are met.

[0048] The beneficial effects of this invention are: through its unique ACSI index and accompanying algorithm, this invention achieves significant technical progress compared to existing methods, and its effects comprehensively surpass traditional empirical combinations and advanced selection algorithms.

[0049] 1. Performance beyond traditional empirical band combinations: In existing remote sensing practices, it is common to rely on "true color" (R-G-B) or "near-infrared-visible" band combinations, or empirical band combinations derived from expert prior experience. In practice, for a specific remote sensing secondary classification task, there is often a lack of prior experience (including the understanding of the spectral physical characteristics of ground objects, and the understanding of the number of band channels and spectral response of remote sensing satellite or unmanned aerial vehicle target sensors), so it is not possible to directly optimize the band combination, and a large amount of time is needed to experimentally verify the optimal band combination. The band combination selected by the ACSI method of the present application is far superior to these traditional empirical band combinations in terms of classification performance. For example, the average intersection over union (mIoU) of the method of the present application is 6.83 percentage points higher than that of the traditional true color combination (R-G-B), and is still 4.2 percentage points higher than that of the combination with the addition of the near-infrared band (R-G-B-NIR), showing the great advantage of the present application in automatically mining deep spectral information.

[0050] 2. The precision is better than the existing advanced algorithm: compared with the mRMR algorithm closest to the current technical idea, the application also shows obvious superiority. From the results, under the same experimental conditions, the mIoU of the ACSI method of the application is 2.11 percentage points higher than that of the mRMR algorithm, the mACC is 2.78% higher, and the mF1 is 1.48% higher. This is due to the unique signal amplification and protection mechanism of ACSI, which successfully retains the key discriminant information that the mRMR correlation prediction measurement mechanism will filter out in the high similarity remote sensing secondary classification. From the principle, mRMR is based on the information theory, measures the statistical dependence of features and class labels through mutual information, and simultaneously constrains the redundancy with a linear weighted balance model, focusing on suppressing the predictability (correlation) at the distribution information level; the application is based on the projection statistical discriminant analysis idea, takes the class separation ratio as the core measurement, establishes a nonlinear power (multiplication) amplification mechanism, so that the weak but key discriminant signal is preferentially retained, and it is also conducive to the amplification of small sample / weak signal (because mRMR does not calculate variance, depends on probability density estimation, and is concerned about the overall optimization); the algorithm realizes the nonlinear decision logic of "signal priority, redundancy subordination" through a multiplication coupling model, which is completely different from the existing linear weighted balance paradigm, and breaks and overcomes the existing algorithm "maximum distinction, minimum redundancy" thinking stereotype and technical bias. The application believes that for remote sensing secondary feature classification, as long as the distinction ability after γ amplification is strong enough, even if its redundancy is high, it does not need to be eliminated, but to retain it; in the highly similar spectral background, ACSI realizes more strict directional exclusion when there are several highly similar features and a key weak signal feature through step-by-step iteration optimization, while the mRMR overall redundancy minimization idea does not give enough punishment to a single signal source, resulting in fewer key signal features actually used for discrimination in the selected feature band. This structural innovation makes ACSI algorithm achieve significant improvement in the highly correlated spectral remote sensing secondary classification task, and the mRMR based on mutual information is different in measurement object and decision logic, which is a non-obvious technical improvement.

[0051] 3. Achieving the efficiency and low cost of "less is more": the method provided by the application can quickly provide the best feature combination, and the mapping task can be completed on single-period multispectral data, which significantly improves the timeliness while ensuring the mapping accuracy. Experiments show that the application can filter out the minimum feature band combination (such as 3-band combination) with the largest amount of effective information, even exceeding the classification effect of full-band (such as 10-band combination), which shows that the effective information filtered by the application can enhance the feature utilization rate of the downstream model without increasing the model complexity, and unexpected results are obtained. The existing mRMR method and other traditional combinations cannot achieve the accuracy of full-band combination (Table 2), which shows that although the traditional method represented by mRMR reduces the feature input, its purpose is subtraction, so the accuracy is not as good as the full-band combination containing all the information. However, the application not only successfully eliminates the confusion information but also purifies the effective information, achieving an effect exceeding the full-band combination. At the same time, this means that the dimension of data processing can be reduced by 70% (in the embodiment) without sacrificing or even improving the accuracy. This directly translates into shorter model training time, lower hardware resource consumption and faster data processing speed, paving the way for large-scale and high-timeliness remote sensing monitoring applications, and has great practical value.

[0052] 4. Strong universality and robustness: the application realizes an adaptive feature selection algorithm, and accordingly realizes a refined mapping framework for remote sensing secondary classification tasks. The core of the application is an adaptive algorithm ACSI based on data-driven and statistical analysis, which does not depend on any prior knowledge or artificial rules. It shows stable high performance on data sets spanning multiple years and different seasons, proving its strong robustness to environmental changes and data fluctuations. Its universality enables it to be seamlessly applied to various refined classification tasks that need to extract and amplify key information from similar spectra.

[0053] In summary, the application fundamentally solves the problem that the prior art cannot effectively retain weak discriminant signals in remote sensing refined classification by creating a new, nonlinear adaptive class separation index ACSI, and provides a technical solution with higher accuracy, better efficiency and stronger universality. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The application is a remote sensing image secondary object refined mapping method based on adaptive discrimination enhancement.

[0055] Figure 2 The application is a remote sensing image secondary object refined mapping method based on adaptive discrimination enhancement.

[0056] Figure 3 The application is a remote sensing image secondary object refined mapping method based on adaptive discrimination enhancement. DETAILED DESCRIPTION

[0057] The application will be further described below in connection with the drawings and specific embodiments.

[0058] As shown in the drawings, Figure 1 A remote sensing image secondary feature fine mapping method based on adaptive discriminant enhancement includes the following steps:

[0059] S1: Select sample points of target secondary classification features and pre-process them to obtain feature category label samples;

[0060] Since the ACSI algorithm is in a supervised form, sample points of target secondary classification features are needed before the algorithm is run, which are used for feature screening.

[0061] The S1 includes the following steps:

[0062] S11: Select sample points according to a preset rule;

[0063] The sample points should be directly provided by coordinate values, and the selection rule of the sample points is that the sample points should be selected at the center of the typical and contiguous area of each feature, the number of sample points of each feature should be >=1, and multiple sample points of the same feature should be evenly distributed as much as possible to cover the entire survey area and the characteristics of the feature.

[0064] S12: Convert the sample points into intersection points with the same resolution of the image grid according to the affine transformation matrix of the remote sensing image to realize satellite pixel grid alignment;

[0065] S13: Take the intersection point as the center and expand outwardly by n grid points (n>=2; n is 3 by default, which can be customized according to the pixel resolution of the remote sensing data used, that is, it should be ensured that the range after 2n x pixel resolution should belong to the coverage range of the feature);

[0066] S14: Calculate the grid mean value for each spectral channel of the image in the expanded grid to obtain the spectral feature curve of the sample point. This sampling strategy eliminates the atypical interference of a single pixel through mean value processing, effectively suppresses the mixed pixel effect at the satellite scale, and then accurately extracts the pure spectral feature curve of different secondary fine-grained classification features.

[0067] S2: Define a spectral band set based on the feature category label samples and initialize an optimal feature set and a candidate feature set;

[0068] The S2 includes the following steps:

[0069] S21: Input remote sensing image data and feature category label samples, define a spectral band set M all ={b1, b2,b3, …, bp}, where p is the total number of bands, and initialize the optimal feature set (Empty set) and the candidate feature set M = M all ;

[0070] S22: Traverse the tunable amplification operator of all candidate bands in the candidate feature set, and select the band with the highest tunable amplification operator value;

[0071] The calculation of the tunable amplification operator includes the following steps:

[0072] S221: For each candidate band in the candidate feature set M , calculate the mean value of each ground object class on the candidate band and the mean value of the sample set on the candidate band ;

[0073] S222: Connect the reflectivity mean values of all sampling sample points of each class of ground objects in each band, i.e. vectorize, to form the spectral feature vector of the class;

[0074] S223: Calculate the Pearson correlation coefficient of the spectral feature vectors of any two classes of ground objects, and calculate the average value of the correlation coefficient of the two combinations according to the class to obtain the overall correlation coefficient ;

[0075] When is large, it means that the similarity is high and the interference is strong, and the difference needs to be nonlinearly enlarged to effectively extract the signal. Therefore, when is used for secondary ground object fine-grained classification task, its value increases exponentially with the increase of the overall correlation coefficient (0<= <=1, =0 means that the ground objects in the task are completely unrelated, and =1 means that the similarity of the ground objects is very high and the difficulty of distinguishing is extremely great). In order to amplify the weak but key discriminant signal in the highly similar spectral background. When < , the spectral similarity of the class has not reached the fine-grained difficulty interval, at this time it belongs to the general ground object classification scene, and does not need to be processed by signal amplification, therefore =1; When >= , it means that the spectral form between classes is highly similar, entering the high interference interval of fine-grained classification, and the nonlinear amplification mechanism based on the separability index needs to be started. To determine whether it belongs to the segmentation point of fine-grained classification, set , this threshold is based on the fact that in statistics A value ≥0.8 is considered a 'strong correlation' and is widely used in remote sensing spectral analysis to identify highly similar spectral curves.

[0076] S224: Determine the adjustable nonlinear amplification index based on the overall correlation coefficient. (1<= <=4, depending on The values ​​are calculated based on the adjustable nonlinear amplification index, the inter-class dispersion, and the intra-class dispersion of the band.

[0077] The value of can be adaptively adjusted based on the average spectral similarity of the secondary land cover categories to be classified; generally speaking, the more similar the spectral curves between different land cover categories are, the more difficult it is to distinguish them, so we can appropriately increase . The value is adjusted to enhance the amplification effect on weak discrimination signals.

[0078] The formula for calculating the adjustable nonlinear amplification index is as follows:

[0079]

[0080]

[0081]

[0082] in, For adjustable nonlinear amplification index, The overall correlation coefficient is... To determine whether it belongs to a segmentation point of fine-grained classification, The number of land cover categories for fine-grained classification. The Pearson correlation coefficient is used. For the first Spectral feature vectors of land-like features For the first Spectral feature vectors of land-like features For the first Land cover samples in characteristic bands The mean of the above, For the first Spectral mean of land-like features across the entire band. For the first Land cover samples in characteristic bands The mean of the above, For the first Spectral mean of land features across the entire band This represents the total number of bands.

[0083] S23: The band b with the highest tunable amplification operator value. bestRemove from the candidate feature set M and add as initial features to the optimal feature set. In the middle, that is, to perform set operations: , ,in, This indicates a union operation, i.e., adding; This represents the difference operation, i.e., removal. At this point, the algorithm initialization is complete, and the optimal set is obtained. It contains an initial feature.

[0084] S3: Based on the optimal feature set and the candidate feature set, determine whether the loop termination condition is met. If not, traverse each band in the candidate feature set, calculate the ACSI value of each band, and store it in the temporary candidate list. Then proceed to step S4. Otherwise, proceed to step S5.

[0085] The S3 step involves traversing each band in the candidate feature set and calculating the ACSI value for each band, including the following sub-steps:

[0086] S31: For each feature band in the candidate feature set Iterate through each selected band in the current optimal feature set. The sample values ​​of each band are concatenated and vectorized to obtain the spectral feature vector, and the feature bands are calculated. With each band Similarity;

[0087] Similarity The calculation formula is:

[0088]

[0089] in, , The first Individual ground feature sample points in the band With band Sample values ​​on The total number of samples.

[0090] S32: Determine the maximum similarity and calculate the characteristic bands. The information gain coefficient relative to the current optimal feature set;

[0091] S33: Calculate the characteristic band based on the information gain coefficient and the tunable amplification operator. ACSI value;

[0092] The application creatively proposes a brand-new feature comprehensive value measurement index, namely, an adaptive class separability index (ACSI), which is specially designed for the field of remote sensing fine-grained classification. The index is unique in its nonlinear and multiplicative coupling mathematical framework, which aims to ensure the absolute dominance of the strongest discriminative signal by the core idea of "first amplifying the key signal, and then modulating it with novelty".

[0093] The ACSI value of the feature band is:

[0094]

[0095] wherein, is the ACSI value of the feature band , is the set of samples of the th land cover type, is the sample number corresponding to the th land cover type, is the mean value of the sample set in the feature band , is the value of any sample in in the feature band , is the set of selected optimal feature bands, and the set length is <= . .

[0096] ACSI is composed of two key parts: the first part is the "tunable amplification operator" of the feature, which is the quantification of the ability of the feature to distinguish different classes, that is, the ratio of the inter-class dispersion to the intra-class dispersion of the feature. The factor is to nonlinearly amplify the small differences in the separability ratio between different features by applying an adjustable nonlinear amplification index . The role of the factor is to nonlinearly amplify the small differences in the separability ratio between different features. The larger the value is, the more significant the amplification effect of the strong signal is, so that the selection of the algorithm for the better feature is more decisive, and the algorithm can be flexibly set according to the similarity of the secondary land cover types to ensure that even a small discriminative advantage can be significantly reflected in the calculation; The value of the total correlation coefficient is exponentially changed, which ensures that is not set by human experience but according to the difficulty of the task (i.e., the total correlation coefficient The second part is the "information gain coefficient" of the feature: it measures the amount of new information that the candidate feature can bring relative to the selected feature set S. This coefficient is determined by calculating the maximum vector direction similarity between the candidate feature and all existing features in the set. The greater the amount of new information, the higher the coefficient value.

[0097] The structure of the formula ensures the unique decision logic of the application when selecting features: the final selection of a feature is overwhelmingly determined by its own nonlinearly amplified class separation ability. Only in the case of similar separation ability, the amount of new information it provides (i.e., low redundancy) becomes a secondary determinant. This "signal absolute priority" strategy is fundamentally different from the "correlation-redundancy" linear trade-off mode of mRMR, perfectly suitable for the high similarity of feature signals in fine classification, and can solve the pain point that key weak signals are easily "killed" by traditional methods (mRMR) in terms of correlation prediction metrics. And as a linear classification calculation formula, mRMR does not have the nonlinear trade-off of the application, nor does it have the adaptive dynamic adjustment capability of the nonlinear proportion of the application, and cannot automatically balance the signal and redundancy in secondary classification tasks of different difficulties.

[0098] The loop termination condition includes:

[0099] (1) The optimal feature set length reaches the preset length;

[0100] The preset length in this embodiment is 3, because most computer vision tasks are prepared for three-channel input, and selecting 3 as the preset value can directly adapt to most existing machine learning models and pre-trained weights without additional configuration and changes. If the user has customization needs, they can also customize the preset feature quantity according to the task accuracy and efficiency balance requirements, but it cannot exceed c.

[0101] (2) The ACSI value of the optimal candidate band is lower than the preset contribution threshold (the contribution threshold preset value is 0.01, because of the existence of the ACSI multiplier effect, the subsequent ACSI value will rapidly decay to achieve improved discrimination).

[0102] The loop terminates when any of the above conditions is met.

[0103] S4: Sort the ACSI values in the temporary candidate list to obtain the band with the highest ACSI value, update the optimal feature set and the candidate feature set using the band, and clear the temporary candidate list. Then return to step S3.

[0104] S5: The cycle is terminated, and the optimal feature set is output in the order of the stack. The bands are input into the machine learning model to generate a secondary feature classification map. This algorithm ensures that each step adds new features that maximize the retention of discriminant signals and bring the most new information.

[0105] In one embodiment of the present application, five types of representative remote sensing secondary classification features are selected in Ya'an, Sichuan, namely "continuous artificial forest, secondary bamboo forest, suitable forest barren mountain, mine, and disaster damaged land". In this combination, "continuous artificial forest", "secondary bamboo forest", and "suitable forest barren mountain" belong to the secondary classification under the first classification "forest land", and other forest land types are unified into one category for comparison; "mine" and "disaster damaged land" belong to the secondary classification under the first classification "bare land", and other bare land types should be classified into one category for comparison. In order to simulate the mixed secondary feature classification requirements often encountered in real business, this embodiment does not separately classify and map the forest or bare land, but actively increases the difficulty to integrate the above five types of secondary features belonging to different first classification into one classification mapping task to fully test the effectiveness of the algorithm. According to the steps of the invention, the team sampled the typical sample points of the above five types of features, and the coordinate values are provided in Table 1. The corresponding satellite images were collected from 2022 to 2024 to simulate different climate conditions and feature phenology changes (equivalent to adding disturbance information, which can test the robustness of the algorithm).

[0106] Table 1: Sampling points

[0107] Number Longitude (East) Latitude (North) Elevation (m) 1 102.898402 29.568048 2142.32 2 102.851339 29.3557532 2463.37 3 102.850225 29.56386 2429.44 4 102.894545 29.604929 1182.65 5 102.815498 29.616804 1595.74 6 102.608184 29.667956 2113.2 7 102.763231 29.625762 1134.97 8 102.55546 29.712042 1226.59 9 102.384700 29.801767 2116.41 10 102.449048 29.757455 1455.98

[0108] After obtaining the sampling points, the sampling points are expanded and sampled, and the pure spectrum curves of the five types of secondary features are obtained, as shown in Figure 2 . Then the average reflectivity of each type of feature is vectorized and connected to obtain the spectral feature vector. The global similarity = 0.9, indicating that the similarity of each feature in this classification task is high, and the classification difficulty will be much greater than that of the first classification task (such as forest, farmland, grassland, water body, and construction land). Figure 2 The curve trend formed by the multiple feature band sequences of = 2.0, which is a square amplification, which is a dynamic signal amplification mechanism that cannot be achieved by existing linear weighting algorithms such as mRMR. This mechanism will effectively amplify the key effective signals in the similar trend ( Figure 2 ) without human intervention.

[0109] Continue to calculate the tunable amplification operator value of all current bands according to the step requirement, and take out the band corresponding to the maximum value (band 12, value 0.563), at this time the first optimal feature has been determined and put into Ω. Then iterative calculation is carried out, and the ACSI value of the selected band in each round of selected band combination with respect to the band in Ω is evaluated, and step-by-step iterative optimization is carried out. In order to show the running results of the algorithm (without termination condition), the output results in the full length Ω are as shown in the figure Figure 3 , in which the length of the band color block indicates the tunable amplification operator, and the filling color indicates the information gain coefficient, and the combination of the two is the ACSI value (specifically: band 12 (0.563), band 4 (0.185), band 8 (0.032), band 11 (0.016), band 2 (0.005), band 5 (0.0025), band 6 (0.0021), band 3 (0.001), band 7 (0.0003), band 8a (0.0001)). In this embodiment, since the default length of Ω is 3, the condition is taken as the stop threshold, so the final output feature band should be: band 12, band 4, and band 8.

[0110] Further, in order to quantify the algorithm effect, we select the closest existing mRMR algorithm and obtain its running results (i.e. band 4, band 8a, band 11), the most commonly used true color combination in computer vision field (i.e. band 4, band 3, band 2), the true color + near-infrared combination recognized in remote sensing field (i.e. band 4, band 3, band 2, band 8), and the full-band combination with full information (i.e. band 2-band 12). Take the four existing technologies or recognized prior knowledge as input, run the same lower classification mapping model with the same hyperparameter combination, and compare the algorithm effect, as shown in Table 2.

[0111] Table 2 Comparison of band combination experiment results

[0112] Method mIoU mAcc mF1 ACSI (Invention) 78.90 90.26 87.78 mRMR (Prior Art) 76.79 87.48 86.30 True Color 72.07 83.56 82.90 True Color + Near Infrared 74.70 85.57 84.89 Full Band Combination 77.15 89.15 86.68

[0113] The comparison of the selected feature bands and the mapping accuracy verification results of the present application and the prior art shows that the selected feature bands of the present application are band 12, band 4, and band 8, and the bands selected by the prior art mRMR (band 4, band 8a, band 11) are only the same as band 4, and the ACSI value of band 8a selected by mRMR is the lowest among all bands (0.0001), which is much lower than the ACSI value of band 8 (0.032) selected by the present application. Figure 3), that is, mRMR and the present application come out completely different results. From the accuracy comparison, the present application achieves the best accuracy in mIoU, mAcc and mF1 three indicators, significantly exceeding the prior art and prior knowledge band (Table 2). Although mRMR exceeds the prior knowledge band, the accuracy cannot exceed the full band combination, because the mRMR method is not optimized for high similarity secondary feature classification, and as an evaluation index with mutual information as the core, it is different from the measurement paradigm and decision logic of the present application with statistical projection analysis as the core. It is worth noting that the accuracy of the present application even exceeds the full band combination, which shows that for remote sensing secondary feature classification mapping tasks, the effective information amount of the input feature three bands selected effectively can reach and even exceed the level of the full band information amount, while eliminating the process of refining effective information from a large amount of interference information, that is, while ensuring the accuracy, the data cost is successfully reduced.

[0114] Those skilled in the art will appreciate that the embodiments described herein are presented for the purpose of helping the reader understand the principles of the present application, and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspiration disclosed in the present application without departing from the essence of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A remote sensing image secondary feature fine mapping method based on adaptive discriminative enhancement, characterized in that, The method comprises the following steps: S1: selecting sample points of target secondary classified ground objects and pre-processing the sample points to obtain ground object category label samples; S2: defining a spectral band set based on the ground object category label samples, and initializing an optimal feature set and a candidate feature set; S3: judging whether a loop termination condition is met based on the optimal feature set and the candidate feature set, if not, traversing each band in the candidate feature set, calculating an ACSI value of each band, and storing the ACSI value in a temporary candidate list, and entering step S4, otherwise entering step S5; The step of traversing each band in the candidate feature set in S3 and calculating an ACSI value of each band comprises the following steps: S31: For each feature band in the candidate feature set , traverse each selected band in the current optimal feature set , concatenate the sample values of each band to obtain a spectral feature vector, and calculate the similarity between the feature band and each band ; S32: Determine the maximum similarity and calculate the characteristic wave band Information gain coefficient relative to the current optimal feature set; S33: calculating the characteristic wave band according to the information gain coefficient and the tunable amplification operator ACSI value; The characteristic wavelength band The ACSI value is: in, For band ACSI value, The number of land cover categories for fine-grained classification. For the first A collection of land cover samples. For the first Number of samples corresponding to land cover types For the first Land cover samples in characteristic bands The mean of the above, For the sample set in the characteristic band The mean of the above, for any sample In characteristic bands The value on, For adjustable nonlinear amplification index, Characteristic bands With each band similarity, This is the set of feature bands that have been selected into the optimal feature set, and the set length is <= ; The adjustable non-linear amplification exponent calculation formula is: wherein is a total correlation coefficient, is a segment point for judging whether it belongs to a fine-grained classification, is a Pearson correlation coefficient, is a spectral feature vector of a first class of ground objects, is a spectral feature vector of a first class of ground objects, is a spectral mean of a first class of ground objects in a full waveband, is a mean of a first class of ground object samples in a feature waveband , is a spectral mean of a first class of ground objects in a full waveband, is a total waveband number; S4: sorting the ACSI values in the temporary candidate list to obtain a band with the highest ACSI value, updating the optimal feature set and the candidate feature set by using the band, and emptying the temporary candidate list, and returning to step S3; S5: terminating the loop, outputting bands in the optimal feature set in a stacking order, and inputting the bands into a machine learning model to generate a secondary ground object refined classification map.

2. The method according to claim 1, wherein, The step S1 comprises the following steps: S11: selecting sample points according to a preset rule; S12: converting the sample points into intersection points with the same resolution as the image grid according to an affine transformation matrix of a remote sensing image; S13: expanding n grid points outward from the intersection point as the center; S14: calculating a grid mean value for each spectral channel of the image in the expanded grid as a spectral feature curve of the sample point. 3.The method of claim 1, wherein the method further comprises: determining a first feature point of the first feature based on the first feature point of the first feature and the first feature point of the second feature; and determining a second feature point of the second feature based on the first feature point of the first feature and the first feature point of the second feature. The step S2 comprises the following steps: S21: inputting remote sensing image data and ground object category label samples, defining a spectral band set, and initializing an optimal feature set and a candidate feature set, the optimal feature set being an empty set; S22: traversing and calculating a tunable amplification operator of all candidate bands in the candidate feature set, and selecting a band with the highest tunable amplification operator value; S23: removing the band with the highest tunable amplification operator value from the candidate feature set and adding the band as an initial feature to the optimal feature set, and completing algorithm initialization.

4. The method according to claim 3, wherein the method is characterized by, The calculation of the tunable amplification operator comprises the following steps: S221: for each candidate band in the candidate feature set, calculating a mean value of each ground object category on the candidate band and a mean value of the sample set on the candidate band; S222: connecting the mean values of all sample sample points of each ground object category on each band to form a spectral feature vector of the category; S223: calculating a Pearson correlation coefficient of the spectral feature vectors of any two categories of ground objects, and calculating an average value of the correlation coefficients of all combinations according to the categories to obtain an overall correlation coefficient; S224: determining an adjustable non-linear amplification exponent according to the overall correlation coefficient, and calculating a tunable amplification operator value based on the adjustable non-linear amplification exponent, an inter-class dispersion of the band, and an intra-class dispersion.

5. The method of claim 1, wherein the method is based on adaptive discriminative enhancement for fine mapping of secondary features in remote sensing imagery. the similarity The calculation formula is: in, , The first Individual ground feature sample points in the band With band Sample values ​​on The total number of samples.

6. The method of claim 1, wherein the method is based on adaptive discriminative enhancement for fine mapping of secondary features in remote sensing imagery. The loop termination condition comprises: (1) the length of the optimal feature set reaches a preset length; (2) the band with the highest ACSI value is recorded as an optimal candidate band, and the ACSI value of the optimal candidate band is lower than a preset contribution threshold. The loop terminates when any of the above conditions are met.

Images