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

Abstract

The invention discloses a remote sensing image secondary ground feature fine drawing method based on adaptive discrimination enhancement, and belongs to the field of remote sensing information processing and geographic information science, and the method comprises the steps: selecting sample points of a target secondary classification ground feature, and carrying out the preprocessing of the sample points; initializing an optimal feature set and a candidate feature set; judging whether a cycle termination condition is met or not, if not, traversing each wave band in the candidate feature set, calculating an ACSI value of each wave band, and storing the ACSI value into a temporary candidate list; sorting the ACSI values in the temporary candidate list to obtain a wave band with the highest ACSI value, and updating the optimal feature set and the candidate feature set by using the wave band; and stopping circulation, outputting wave bands in the optimal feature set according to a stack entering sequence, inputting each wave band into a machine learning model, and generating a secondary surface feature refined classification map. According to the method, the problem that weak discrimination signals cannot be effectively reserved in remote sensing fine classification in the prior art is fundamentally solved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing information processing and geographic information science, specifically relating to a method for refining secondary ground features in remote sensing images based on adaptive discrimination enhancement. Background Technology

[0002] In applications such as ecological monitoring, precision agriculture, and urban planning, accurately distinguishing between secondary land cover categories with highly similar spectral characteristics (e.g., differentiating between coniferous forests and broad-leaved forests, rather than primary land cover such as vegetation, water bodies, and cities) has long been a technical bottleneck. When weather conditions (such as clouds and rain) prevent the acquisition of continuous time-series imagery, or when it is impossible to use expensive multi-source data (such as synthetic aperture radar and high spatial resolution optical data), the ability to efficiently extract the most discriminative information from single-temporal multispectral imagery of sparsely datad areas becomes crucial to the success of the mission and a key factor in determining whether timely periodic operational monitoring can be guaranteed.

[0003] Existing single-temporal image feature selection methods primarily aim to balance the goals of maximizing class discriminative power and minimizing information redundancy. Among these, the most representative and technologically similar advanced method is the Minimum Redundancy Maximum Relevance (mRMR) algorithm. Based on the mutual information criterion, the mRMR algorithm attempts to simultaneously optimize two objectives: maximizing the correlation between each feature (band) and the final classification label, and minimizing the redundancy between features. It has wide applications in bioinformatics and remote sensing primary classification tasks. However, when handling refined remote sensing secondary classification tasks, mRMR and similar methods reveal significant shortcomings: 1. Over-penalization of weak signals: The essence of refined classification is often to capture key signals composed of slight differences in reflectance values ​​in specific bands within a highly correlated spectral background. The "correlation-redundancy" balancing mechanism of mRMR imposes excessive penalties on bands with similar spectral trends (i.e., high redundancy). This can easily lead to the erroneous filtering out of bands that, while correlated, contain crucial discriminative information, resulting in the permanent loss of effective information.

[0004] 2. Limitations of the 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 the signal and the harm of redundancy are of equal magnitude. However, in refined classification, since they belong to the same large category, their spectral reflectance curves will be similar in waveform. What reflects their differences is often the value of a weak but unique discriminative signal, which may far outweigh the negative impact of the redundancy it brings.

[0005] 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.

[0006] 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

[0007] 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.

[0008] 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: S1: Select sample points of the target secondary classification land features and preprocess them to obtain land feature category label samples; S2: Based on the land cover category label samples, define the spectral band set and initialize the optimal feature set and candidate feature set; 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. 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. 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.

[0009] Further, step S1 includes the following sub-steps: S11: Select sample points according to preset rules; S12: Convert the sample points into intersection points with the same resolution as the image grid according to the affine transformation matrix of the remote sensing image; S13: Expand outward from the intersection point by n grid points; S14: Within the expanded grid, calculate the grid mean for each spectral channel of the image, which serves as the spectral characteristic curve for that sample point.

[0010] Furthermore, step S2 includes the following sub-steps: S21: Input remote sensing image data and land cover category label samples, define a spectral band set, and initialize the optimal feature set and candidate feature set, wherein the optimal feature set is an empty set; S22: Iterate through and calculate the tunable amplification operator for all candidate bands in the candidate feature set, and select the band with the highest tunable amplification operator value. S23: Remove the band with the highest tunable amplification operator value from the candidate feature set and add it as an initial feature to the optimal feature set to complete the algorithm initialization.

[0011] Furthermore, the calculation of the tunable amplification operator includes the following sub-steps: S221: For each candidate band in the candidate feature set, calculate the mean of each land cover category on the candidate band and the mean of the sample set on the candidate band; S222: Connect the mean values ​​of all sample points of each land cover type in each band to form the spectral feature vector of that type; S223: Calculate the Pearson correlation coefficient of the spectral feature vectors of any two land cover types, and calculate the average correlation coefficient of each pair of combinations according to the categories to obtain the overall correlation coefficient; S224: Determine the adjustable nonlinear amplification index based on the overall correlation coefficient, and calculate the tunable amplification operator value based on the adjustable nonlinear amplification index, the inter-class dispersion and intra-class dispersion of the band.

[0012] Furthermore, the formula for calculating the adjustable nonlinear amplification index is as follows: 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 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.

[0013] 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: 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; S32: Determine the maximum similarity and calculate the characteristic bands. The information gain coefficient relative to the current optimal feature set; S33: Calculate the characteristic band based on the information gain coefficient and the tunable amplification operator. ACSI value.

[0014] Furthermore, 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.

[0015] Furthermore, the characteristic band The ACSI value is: 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 <= .

[0016] Furthermore, the loop termination condition includes: (1) The length of the optimal feature set reaches the preset length; (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. The loop terminates when any of the above conditions are met.

[0017] 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. 1. Performance Surpassing Traditional Empirical Combinations: In existing remote sensing practices, reliance is typically placed on "true color" (RGB) or "near-infrared-visible light" band combinations, or empirical band combinations derived from prior expert experience. In practice, for a specific remote sensing secondary classification task, prior experience is often lacking (including understanding of the spectral physical characteristics of ground objects, and the number of band channels and spectral response of remote sensing satellites or UAV target sensors), making it impossible to directly optimize band combinations and requiring significant time for experimental verification of the best combinations. The ACSI method of this invention adaptively selects band combinations whose classification performance far surpasses these traditional empirical combinations. For example, the average intersection-union ratio (mIoU) of the method of this invention is 6.83 percentage points higher than that of the traditional true color combination (RGB), and still 4.2 percentage points higher than the combination that adds the near-infrared band (RGB-NIR), demonstrating the significant advantage of this invention in automatically mining depth spectral information.

[0018] 2. Superior Accuracy Compared to Existing Advanced Algorithms: Compared to the mRMR algorithm, which has the closest technical approach, this invention also demonstrates significant superiority. Results show that, under the same experimental conditions, the ACSI method of this invention achieves a 2.11 percentage point higher mIoU, a 2.78% higher mACC, and a 1.48% higher mF1 than the mRMR algorithm. This is attributed to ACSI's unique signal amplification and protection mechanism, which successfully preserves the crucial discriminative information that the mRMR correlation prediction metric would filter out in highly similar remote sensing secondary classification. In principle, mRMR is based on information theory, measuring the statistical dependence of features and class labels through mutual information, and simultaneously constraining redundancy with a linear weighted balancing model, focusing on suppressing predictability (correlation) at the distribution information level. This invention, however, is based on projective statistical discriminant analysis, using the class separation ratio as the core metric, and establishing a nonlinear power (multiplicative) amplification mechanism. This prioritizes the preservation of weak but crucial discriminative signals, while also amplifying small samples / weak signals (because mRMR does not calculate variance, relies on probability density estimation, and focuses on overall optimization). This algorithm achieves a nonlinear decision-making process of "signal priority, redundancy subordination" through a multiplicative coupling model. The logic of this invention is completely different from the existing linear weighted balance paradigm. It also breaks and overcomes the mindset and technical bias of existing algorithms that prioritize "maximum discrimination and minimum redundancy." This invention argues that for remote sensing secondary land cover classification, as long as the discrimination ability after γ-amplification is strong enough, even if its redundancy is high, it does not need to be eliminated but rather retained. In highly similar spectral backgrounds, ACSI achieves stricter directional exclusion when several highly similar features and a key weak signal feature exist through iterative optimization. In contrast, the idea of ​​minimizing overall redundancy in mRMR does not sufficiently penalize a single signal source, resulting in fewer key signal features actually used for discrimination in the selected feature bands. This structural innovation enables the ACSI algorithm to achieve significant improvements in spectrally highly correlated remote sensing secondary classification tasks. It differs from the mutual information-based mRMR in both its measurement object and decision logic, representing a non-obvious technical improvement.

[0019] 3. Achieving high efficiency and low cost with "less is more": The method provided by this invention can quickly provide the optimal feature combination and complete the mapping task on single-period multispectral data, significantly improving timeliness while ensuring mapping accuracy. Experiments show that this invention can filter out the smallest feature band combination with the largest amount of effective information (such as a 3-band combination) and even surpass the classification effect of the full band (such as a 10-band combination). This indicates that the effective information after ACSI filtering can enhance the feature utilization rate of the downstream model without increasing model complexity, resulting in unexpected results. Existing mRMR methods and other traditional combinations cannot achieve the accuracy of the full band combination (Table 2). This shows that although traditional methods represented by mRMR reduce feature input, their purpose is subtraction, so it is predictable that the accuracy is not as good as the full band combination containing all information. However, this invention successfully removes obfuscated information while purifying effective information, achieving an effect that surpasses the full band combination. At the same time, this means that the dimensionality of data processing can be reduced by 70% without sacrificing or even improving accuracy (in the examples). This directly translates to shorter model training time, lower hardware resource consumption, and faster data processing speed, paving the way for large-scale, high-time-efficiency remote sensing monitoring applications, and is of great practical value.

[0020] 4. Strong versatility and robustness: This invention implements an adaptive feature selection algorithm and based on it, a framework for refined mapping of remote sensing secondary classification tasks. The core of this invention is an adaptive algorithm, ACSI, entirely based on data-driven and statistical analysis, which does not rely on any prior knowledge or manual rules. It exhibits stable high performance on datasets spanning multiple years and different seasons, demonstrating its strong robustness to environmental changes and data fluctuations. Its versatility allows it to be seamlessly applied to various refined classification tasks that require extracting and amplifying key information from similar spectra.

[0021] In summary, this invention fundamentally solves the problem that existing technologies cannot effectively preserve weak discrimination signals in remote sensing fine classification by creating a novel, nonlinear adaptive class separation index (ACSI), and provides a technical solution with higher accuracy, better efficiency, and greater versatility. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method for refining secondary ground features in remote sensing images based on adaptive discriminant enhancement, as described in this invention.

[0023] Figure 2 This is a comparison chart of the pure spectral curves of five types of land features in Example 1.

[0024] Figure 3 This is a graph showing the results of running ACSI. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0026] like Figure 1 As shown, a method for refining secondary land cover mapping based on adaptive discriminative enhancement of remote sensing images includes the following steps: S1: Select sample points of the target secondary classification land features and preprocess them to obtain land feature category label samples; Since the ACSI algorithm is supervised, it requires sample points of the target secondary classification land cover to be provided before the algorithm runs for feature selection.

[0027] S1 includes the following steps: S11: Select sample points according to preset rules; Sample points should be provided directly by coordinate values. The selection rules for sample points are as follows: sample points should be selected in the center of typical and contiguous areas of each type of land cover. The number of sample points for each type of land cover should be >= 1. Multiple sample points of the same type of land cover should be distributed as evenly as possible to cover the entire survey area and the characteristics of that type of land cover.

[0028] S12: Convert the sample points into intersection points with the same resolution as the image grid according to the affine transformation matrix of the remote sensing image, so as to achieve satellite pixel grid alignment; S13: Expand outward from the intersection point by n grid points (n>=2; n defaults to 3, which can be customized according to the pixel resolution of the remote sensing data used, i.e., the range after 2n×pixel resolution should be within the coverage area of ​​this type of land cover). S14: Within the expanded grid, calculate the grid mean for each spectral channel of the image, which serves as the spectral characteristic curve for that sample point. This sampling strategy eliminates atypical interference from individual pixels through mean normalization, while effectively suppressing the mixed pixel effect at the satellite scale, thereby accurately extracting the pure spectral characteristic curves of different secondary fine-grained land cover classifications.

[0029] S2: Based on the land cover category label samples, define the spectral band set and initialize the optimal feature set and candidate feature set; S2 includes the following steps: S21: Input remote sensing image data and land cover category label samples, and define the spectral band set M. all ={b1, b2, b3, …, b p}, where p is the total number of bands, and the optimal feature set is initialized. (Empty set) and candidate feature set M=M all ; S22: Iterate through and calculate the tunable amplification operator for all candidate bands in the candidate feature set, and select the band with the highest tunable amplification operator value. The calculation of the tunable amplification operator includes the following steps: S221: For each candidate band in the candidate feature set M Calculate the categories of objects in various regions mean on the candidate band and the mean of the sample set in the candidate band ; S222: Connect the average reflectance values ​​of all sample points of each land cover type in each band, i.e. vectorize them, to form the spectral feature vector of that type. S223: Calculate the Pearson correlation coefficient between the spectral feature vectors of any two land cover types, and calculate the average correlation coefficient of each pairwise combination according to the category to obtain the overall correlation coefficient. ; when A larger value indicates higher similarity and stronger interference, requiring nonlinear scaling to increase the difference for effective signal extraction. Therefore, when... When used for secondary fine-grained land cover classification tasks, its value varies with the overall correlation coefficient. (0<= <=1, =0 indicates that the ground features in the task are completely unrelated, while =1 indicates that the ground features are very similar and extremely difficult to distinguish. The value increases exponentially to amplify weak but crucial discrimination signals in a highly similar spectral background. When < Since the spectral similarity of the categories has not yet reached the fine-grained difficulty range, this belongs to a general land cover classification scenario, and no signal amplification processing is required. Therefore, we set... =1; when >= If the spectral morphology of the categories is highly similar, it indicates that the categories have entered a high-interference range for fine-grained classification, and a nonlinear amplification mechanism based on the separability index needs to be activated. To determine whether a segment point belongs to the fine-grained classification, set... This threshold is based on the statistically common practice of... A value ≥0.8 is considered a 'strong correlation' and is widely used in remote sensing spectral analysis to identify highly similar spectral curves.

[0030] 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. 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.

[0031] The formula for calculating the adjustable nonlinear amplification index is as follows: 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 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.

[0032] S23: The band b with the highest tunable amplification operator value. best Remove 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.

[0033] 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. 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: 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; 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.

[0034] S32: Determine the maximum similarity and calculate the characteristic bands. The information gain coefficient relative to the current optimal feature set; S33: Calculate the characteristic band based on the information gain coefficient and the tunable amplification operator. ACSI value; This invention creatively proposes a novel feature comprehensive value metric specifically for the field of fine-grained classification in remote sensing—the Adaptive Class Separability Index (ACSI). The unique feature of this index lies in its non-linear, multiplicative-coupled mathematical architecture, which, through the core idea of ​​"first amplifying the key signal, then modulating it with novelty," aims to prioritize ensuring the absolute dominance of the strongest discriminative signal.

[0035] The characteristic band The ACSI value is: 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 <= .

[0036] ACSI consists of two key parts: the first part is the feature's "tunable amplification operator," which quantifies the feature's ability to distinguish between different classes, i.e., it measures the ratio of the feature's inter-class dispersion to its intra-class dispersion. This factor is achieved by applying an adjustable non-linear amplification exponent. , Its function is to non-linearly amplify the small differences in the separability ratio between different features. The larger the value, the more significant the amplification effect on strong signals, making the algorithm more decisive in selecting better features. It can be flexibly set adaptively according to the similarity of secondary land cover categories to ensure that even a small discrimination advantage can be significantly reflected in the calculation. Correlation coefficient between the value and the population It changes exponentially, which ensures It is not set by human experience, but based on the difficulty of the task (i.e., the overall correlation coefficient). The first part is the dynamic calculation of the size of the feature. 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 value of this coefficient.

[0037] The structure of this formula ensures the unique decision-making logic of this invention when selecting features: the final selection qualification of a feature overwhelmingly depends on its class separation ability after nonlinear amplification. Only when the separation abilities are similar will the amount of new information it provides (i.e., low redundancy) become a secondary determining factor. This "signal absolute priority" strategy is fundamentally different from the "correlation-redundancy" linear trade-off mode of mRMR, perfectly suited to the high similarity of ground object signals in refined classification, and can solve the problem that key weak signals are easily "falsely eliminated" by traditional methods (mRMR) based on correlation prediction. Moreover, as a linear classification calculation formula, mRMR does not have nonlinear trade-offs, nor does it have the nonlinear proportions required in this invention. Its adaptive dynamic adjustment capability is insufficient to achieve an automatic trade-off between signal and redundancy in secondary classification tasks of varying difficulty.

[0038] The loop termination condition includes: (1) The length of the optimal feature set reaches the preset length; In this embodiment, the preset length is 3, because most computer vision tasks are prepared for three-channel input. Choosing 3 as the preset value can directly adapt to most existing machine learning models and pre-trained weights without additional configuration or changes. If the user has customization needs, the preset number of features can also be customized to meet the balance between task accuracy and efficiency, but it cannot exceed c.

[0039] (2) The band with the highest ACSI value is recorded as the optimal candidate band. The ACSI value of the optimal candidate band is lower than the preset contribution threshold (the preset contribution threshold is 0.01, because the existence of the ACSI multiplier effect will cause the subsequent ACSI value to decay rapidly in order to improve the distinguishability). The loop terminates when any of the above conditions are met.

[0040] 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. 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 refinement map. This algorithm ensures that each new feature added at each step is the optimal choice that brings the most new information while maximizing the preservation of the discriminative signal.

[0041] In one embodiment of the present invention, five representative types of remote sensing secondary land cover were selected in Ya'an, Sichuan Province: "contiguous artificial forests, secondary bamboo forests, suitable barren hills for afforestation, mines, and disaster-damaged land." In this combination, "contiguous artificial forests," "secondary bamboo forests," and "suitable barren hills for afforestation" belong to the secondary land cover category under the primary category "forest land," while other forest land types are grouped into one category for comparison. "Mine land" and "disaster-damaged land" both belong to the secondary land cover category under the primary category "bare land," while other bare land types should be grouped into one category for comparison. To simulate the mixed secondary land cover classification requirements often encountered in real-world operations, this embodiment does not separately classify and map forest land or bare land. Instead, it proactively increases the difficulty by integrating the above five secondary land cover categories belonging to different primary land cover categories into a single classification and mapping task to fully test the effectiveness of the algorithm. Following the steps required by the invention, the team sampled typical points of the above five types of land cover, and the coordinate values ​​are provided in Table 1. The corresponding satellite images were collected between 2022 and 2024 to simulate different climate conditions and changes in land cover and phenology (equivalent to adding perturbation information, which can test the robustness of the algorithm).

[0042] Table 1 Sampling Points serial number Longitude (East) Latitude (North Latitude) Altitude (meters) 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 After obtaining the sampling points, expanded sampling was carried out using these points as the center, and the pure spectral curves of five types of secondary land cover were obtained, such as... Figure 2 As shown. Then, the mean reflectance values ​​of each land cover type are vectorized and concatenated to obtain the spectral feature vector. Global similarity is then calculated. =0.9 indicates that the similarity of features in different regions is high in this classification task, and the classification difficulty will be much greater than that of the first-level classification task (such as forest land, cultivated land, grassland, water body, construction land, etc.). Figure 2 The trend and degree of mixing of the curve formed by multiple characteristic band sequences can also intuitively indicate this conclusion. At this point, the algorithm will automatically amplify the class separation index according to the difficulty of classification, and the calculation will yield... =2.0, belonging to quadratic amplification, which is a dynamic signal amplification mechanism that existing linear tradeoff algorithms such as mRMR cannot achieve. This mechanism will effectively amplify similar trends ( Figure 2 It provides key and effective signals without the need for manual parameter adjustment.

[0043] Continue calculating the tunable amplification operator values ​​for all current bands as required, and extract the band corresponding to the maximum value (band 12, value 0.563). At this point, the first optimal feature has been determined and placed into Ω. Then, iterative calculations are performed, and the ACSI values ​​of the candidate bands in each round of candidate band combinations relative to the bands in Ω are evaluated, and progressive iterative optimization is carried out. To demonstrate the continuous operation of the algorithm (without a termination condition), the output results in the full-length Ω are as follows: Figure 3 As shown in the figure, the length of the band color block indicates the tunable amplification operator, while the fill color indicates the information gain coefficient. 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, this condition is used as the stopping threshold, so the final output characteristic bands should be: band 12, band 4, and band 8.

[0044] Furthermore, to quantify the algorithm's performance, we selected the closest existing mRMR algorithm and obtained its results (i.e., bands 4, 8a, and 11), the most commonly used true-color combination in computer vision (i.e., bands 4, 3, and 2), the universally recognized true-color + near-infrared combination in remote sensing (i.e., bands 4, 3, 2, and 8), and the full-band combination with complete information (i.e., bands 2-12). Using these four existing technologies or universally recognized prior knowledge as input, we ran the same downstream classification and mapping model with the same hyperparameter combinations to compare the algorithm's performance. The results are shown in Table 2.

[0045] Table 2 Comparison of Band Combination Experiment Results method mIoU mAcc mF1 ACSI (This Invention) 78.90 90.26 87.78 mRMR (existing technology) 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 Comparison of the characteristic bands selected by the present invention and the mapping accuracy verification results with those of the prior art shows that the characteristic bands selected by the present invention 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 identical to band 4. Moreover, the ACSI value of band 8a selected by mRMR is the lowest among all bands. Figure 3 This means that the mRMR and the present invention yielded completely different results. In terms of accuracy comparison, the present invention achieved the best accuracy in all three metrics: mIoU, mAcc, and mF1, significantly exceeding existing technologies and prior knowledge bands (Table 2). Although the mRMR exceeded that of the prior knowledge bands, its accuracy did not surpass that of the full band combination. This is because the mRMR method was not optimized for secondary land cover classification with high similarity, and as an evaluation index based on mutual information, it differs from the measurement paradigm and decision logic of the present invention, which is based on statistical projection analysis. It is noteworthy that the accuracy of the present invention even surpasses that of the full band combination. This indicates that for remote sensing secondary land cover classification and mapping tasks, the effective information content of the effectively selected input feature three bands can reach and exceed the level of the full band information content, while eliminating the process of extracting effective information from a large amount of interference information. In other words, it successfully reduces data costs while ensuring accuracy.

[0046] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the invention.

Claims

1. A method for refining secondary ground cover mapping of remote sensing images based on adaptive discriminative enhancement, characterized in that, Includes the following steps: S1: Select sample points of the target secondary classification land features and preprocess them to obtain land feature category label samples; S2: Based on the land cover category label samples, define the spectral band set and initialize the optimal feature set and candidate feature set; 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. 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. 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.

2. The method for refining secondary ground cover mapping of remote sensing images based on adaptive discriminant enhancement according to claim 1, characterized in that, S1 includes the following steps: S11: Select sample points according to preset rules; S12: Convert the sample points into intersection points with the same resolution as the image grid according to the affine transformation matrix of the remote sensing image; S13: Expand outward from the intersection point by n grid points; S14: Within the expanded grid, calculate the grid mean for each spectral channel of the image, which serves as the spectral characteristic curve for that sample point.

3. The method for refined mapping of secondary ground features in remote sensing images based on adaptive discriminant enhancement according to claim 1, characterized in that, S2 includes the following steps: S21: Input remote sensing image data and land cover category label samples, define a spectral band set, and initialize the optimal feature set and candidate feature set, wherein the optimal feature set is an empty set; S22: Iterate through and calculate the tunable amplification operator for all candidate bands in the candidate feature set, and select the band with the highest tunable amplification operator value. S23: Remove the band with the highest tunable amplification operator value from the candidate feature set and add it as an initial feature to the optimal feature set to complete the algorithm initialization.

4. The method for refined mapping of secondary ground features in remote sensing images based on adaptive discriminant enhancement according to claim 3, characterized in that, The calculation of the tunable amplification operator includes the following steps: S221: For each candidate band in the candidate feature set, calculate the mean of each land cover category on the candidate band and the mean of the sample set on the candidate band; S222: Connect the mean values ​​of all sample points of each land cover type in each band to form the spectral feature vector of that type; S223: Calculate the Pearson correlation coefficient of the spectral feature vectors of any two land cover types, and calculate the average correlation coefficient of each pair of combinations according to the categories to obtain the overall correlation coefficient; S224: Determine the adjustable nonlinear amplification index based on the overall correlation coefficient, and calculate the tunable amplification operator value based on the adjustable nonlinear amplification index, the inter-class dispersion and intra-class dispersion of the band.

5. The method for refining secondary ground cover mapping of remote sensing images based on adaptive discriminant enhancement according to claim 4, characterized in that, The formula for calculating the adjustable nonlinear amplification index is as follows: 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-like features across the entire band. This represents the total number of bands.

6. The method for refined mapping of secondary ground features in remote sensing images based on adaptive discriminant enhancement according to claim 5, characterized in that, 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: 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; S32: Determine the maximum similarity and calculate the characteristic bands. The information gain coefficient relative to the current optimal feature set; S33: Calculate the characteristic band based on the information gain coefficient and the tunable amplification operator. ACSI value.

7. The method for refined mapping of secondary ground features in remote sensing images based on adaptive discriminant enhancement according to claim 6, characterized in that, 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.

8. The method for refining secondary ground cover mapping of remote sensing images based on adaptive discriminant enhancement according to claim 7, characterized in that, The characteristic band The ACSI value is: 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 <= .

9. The method for refined mapping of secondary ground features in remote sensing images based on adaptive discriminant enhancement according to claim 1, characterized in that, The loop termination condition includes: (1) The length of the optimal feature set reaches the preset length; (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 the preset contribution threshold. The loop terminates when any of the above conditions are met.

Images