A method and system for identifying natural propagation and artificial planting of mangrove forests

Abstract

The application discloses a natural propagation and artificial planting mangrove forest identification method and system, relates to the technical field of data processing, and comprises the following steps: extracting a comprehensive feature matrix of a research area in a target year; determining a plurality of initial sampling points in a known artificial forest patch; extracting an initial feature matrix of the initial sampling points from the comprehensive feature matrix; determining a target sampling point in a newly added mangrove forest area according to the initial feature matrix; extracting a target feature matrix of the target sampling point from the comprehensive feature matrix; inputting the target feature matrix into a pre-trained ensemble learning model to obtain a comprehensive probability distribution; and dividing the newly added mangrove forest area into at least one of natural propagation, artificial planting or natural expansion after human intervention according to the comprehensive probability distribution. The application can learn from the sampling points of the known artificial forest as a small sample, accurately divide the newly added mangrove forest area, and finally identify the type of the newly added mangrove forest area.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for identifying naturally propagated and artificially planted mangroves. Background Technology

[0002] Mangroves are unique woody plant communities growing in the intertidal zone of tropical and subtropical coasts. As one of the world's most valuable natural assets for ecosystem services, they are not only key carriers of "blue carbon sinks" for addressing climate change—their carbon storage per unit area can be 3-5 times that of tropical rainforests—but also natural ecological barriers along the coast, playing an irreplaceable role in mitigating typhoon storm surge energy, maintaining coastal biodiversity, and supporting the livelihoods of coastal communities. Large-scale, systematic mangrove protection and restoration are underway globally. Against this backdrop, accurately assessing restoration effectiveness and calculating carbon sequestration increases has become a core task with both significant scientific value and urgent management needs.

[0003] In restoration practice, the sources of newly added mangrove forests are complex, mainly involving three mechanisms: areas of direct human intervention (formed through artificial planting, possessing a regular spatial pattern, and serving as the direct basis for carbon sequestration project verification and trading), areas of natural expansion after human intervention (using the former as the seed source, formed through natural seed dispersal and settlement, reflecting the ecological radiation effect of artificial restoration), and areas of purely natural growth (formed through natural succession without direct intervention, reflecting the ecosystem's baseline restoration capacity). Therefore, scientifically distinguishing these three types of areas is not only a prerequisite for quantifying the direct benefits of artificial afforestation, but also a key basis for assessing natural restoration potential, optimizing spatial restoration layout, and formulating differentiated management strategies. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method and system for identifying naturally propagated and artificially planted mangroves, so as to accurately identify the type of newly added mangroves.

[0005] One aspect of this application provides a method for identifying naturally propagated and artificially planted mangroves, the method comprising the following steps:

[0006] Extract the comprehensive feature matrix of the study area within the target year;

[0007] Multiple initial sampling points were identified within known plantation patches in the study area;

[0008] Extract the initial feature matrix corresponding to each initial sampling point from the comprehensive feature matrix;

[0009] Based on the initial feature matrix, multiple target sampling points are determined within the newly added mangrove area in the study area;

[0010] Extract the target feature matrix corresponding to each target sampling point from the comprehensive feature matrix;

[0011] Each of the target feature matrices is input into a pre-trained ensemble learning model to obtain the comprehensive probability distribution output by the pre-trained ensemble learning model.

[0012] Based on the comprehensive probability distribution, the newly added mangrove area is classified into at least one of the following: natural propagation, artificial planting, or natural expansion after human intervention.

[0013] In some embodiments, extracting the comprehensive feature matrix of the study area within the target year includes the following steps:

[0014] The spectral features of the study area in the target year, the structural features measured by synthetic aperture radar, the topographic background features, the temporal harmonic features, and the multi-scale texture features are extracted as the comprehensive feature matrix.

[0015] In some embodiments, determining multiple target sampling points within newly added mangrove areas in the study area based on the initial feature matrix includes the following steps:

[0016] For each unlabeled pixel in the newly added mangrove area, the cosine similarity between the corresponding feature vector and the initial feature matrix of each initial sampling point is calculated, and the maximum value is taken; wherein, the comprehensive feature matrix includes the feature vector of the unlabeled pixel;

[0017] If the cosine similarity of the maximum value is greater than the first set threshold, then the corresponding unlabeled pixel is determined as the target sampling point for the expanded positive sample;

[0018] If the cosine similarity of the maximum value is less than the second set threshold, then the corresponding unlabeled pixel is determined as the target sampling point of the confidence negative sample; wherein, the second set threshold is less than the first set threshold.

[0019] In some embodiments, the ensemble learning model includes an XGBoost classifier and a random forest classifier;

[0020] The pre-trained ensemble learning model is obtained through the following steps:

[0021] The XGBoost classifier and the random forest classifier are trained in parallel; both classifiers use five-fold cross-validation to optimize hyperparameters.

[0022] Based on the weighted F1 scores on the validation set, the optimal ensemble weights for the XGBoost classifier and the random forest classifier are dynamically calculated respectively.

[0023] The probability distributions output by the XGBoost classifier and the random forest classifier are weighted and summed according to their respective optimal ensemble weights to obtain the comprehensive probability distribution; wherein, the probability distribution represents the probability that each unlabeled pixel belongs to a human-intervention patch.

[0024] In some embodiments, the comprehensive probability distribution is calculated using the following expression:

[0025] ;

[0026] in, For the aforementioned comprehensive probability distribution, The optimal ensemble weights for the XGBoost classifier are... The probability distribution output by the XGBoost classifier is... The optimal ensemble weights for the random forest classifier are... This represents the probability distribution output by the random forest classifier.

[0027] In some embodiments, classifying the newly added mangrove area according to the comprehensive probability distribution into at least one of natural propagation, artificial planting, or natural expansion after human intervention includes the following steps:

[0028] The pixel coordinates classified as human intervention are selected based on the comprehensive probability distribution.

[0029] Several candidate artificial forest patches are determined based on the distance between the coordinates of the pixels involved in each human intervention.

[0030] If the continuous area of ​​the candidate plantation patch reaches a preset area threshold, it is determined to be an artificially planted patch.

[0031] If the continuous area of ​​the candidate plantation patch is less than the preset area threshold, then the minimum Euclidean distance to the centroid of each plantation patch is calculated.

[0032] If the minimum Euclidean distance is less than or equal to a preset spatial proximity threshold, then the corresponding candidate plantation patch is determined to be a patch that has naturally expanded after human intervention.

[0033] If the minimum Euclidean distance is greater than the preset spatial proximity threshold, then the corresponding candidate plantation patch is determined to be a naturally proliferating patch.

[0034] In some embodiments, determining a plurality of candidate artificial forest patches based on the distance between the coordinates of the individual human interventions includes the following steps:

[0035] The DBSCAN algorithm is used to cluster the coordinates of each human intervention cell and remove noise coordinates to obtain several non-noise clusters. Each non-noise cluster is a candidate artificial forest patch.

[0036] Another aspect of this application embodiment provides a system for identifying naturally propagated and artificially planted mangroves, the system comprising:

[0037] The comprehensive feature extraction unit is used to extract the comprehensive feature matrix of the study area within the target year;

[0038] The first anchor point unit is used to determine multiple initial sampling points within known plantation patches in the study area;

[0039] An initial feature extraction unit is used to extract the initial feature matrix corresponding to each initial sampling point from the comprehensive feature matrix;

[0040] The second anchor point unit is used to determine multiple target sampling points within the newly added mangrove area in the study area based on the initial feature matrix;

[0041] The target feature extraction unit is used to extract the target feature matrix corresponding to each target sampling point from the comprehensive feature matrix;

[0042] The probability prediction unit is used to input each of the target feature matrices into the pre-trained ensemble learning model to obtain the comprehensive probability distribution output by the pre-trained ensemble learning model.

[0043] The regional division unit is used to divide the newly added mangrove area into at least one of the following categories based on the comprehensive probability distribution: natural propagation, artificial planting, or natural expansion after human intervention.

[0044] Another aspect of this application embodiment provides an electronic device, including a processor and a memory;

[0045] The memory is used to store programs;

[0046] The processor executes the program to implement any of the methods described above.

[0047] Another aspect of this application provides a computer-readable storage medium storing a program that is executed by a processor to implement the method described in any of the above embodiments.

[0048] This application includes at least the following beneficial effects:

[0049] This application can extract a comprehensive feature matrix of a study area in a target year; determine multiple initial sampling points within known plantation patches in the study area; extract initial feature matrices corresponding to each initial sampling point from the comprehensive feature matrix; determine multiple target sampling points within newly added mangrove areas in the study area based on the initial feature matrices; extract target feature matrices corresponding to each target sampling point from the comprehensive feature matrix; input each target feature matrix into a pre-trained ensemble learning model to obtain a comprehensive probability distribution output by the pre-trained ensemble learning model; and classify the newly added mangrove areas into at least one of natural propagation, artificial planting, or natural expansion after human intervention based on the comprehensive probability distribution. This application can accurately classify newly added mangrove areas by using sampling points from known plantations as small samples for learning, and ultimately identify the type of newly added mangrove areas. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart illustrating a method for identifying naturally propagated and artificially planted mangroves, provided in an embodiment of this application;

[0052] Figure 2 An example flowchart illustrating a method for identifying naturally propagated and artificially planted mangroves, provided as an embodiment of this application;

[0053] Figure 3 This is a structural block diagram of a system for identifying naturally propagated and artificially planted mangroves, provided in an embodiment of this application. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0055] Before providing a detailed description of the embodiments of this application, some related technologies involved in the embodiments of this application will be described first, as follows:

[0056] Terminology Explanation:

[0057] Table 1

[0058]

[0059] The relevant technologies are as follows:

[0060] Currently, remote sensing technology has become an irreplaceable means of achieving large-scale mangrove monitoring. From a technical perspective, mangrove growth type identification is a typical high-dimensional feature classification problem under multi-source information fusion. An ideal solution requires the coordinated processing of spectral information from optical images, structural information from SAR data, phenological information from time-series images, and background information from topography. However, existing technologies still face a series of interrelated fundamental challenges when facing detailed identification tasks: First, the extreme scarcity of training samples constitutes a fundamental methodological obstacle. Supervised learning, as the mainstream paradigm, highly depends on a large number of high-quality labeled samples for its performance. However, obtaining mangrove plantation samples is extremely costly, requiring confirmation through historical archives, high-resolution image retrospection, and field verification, resulting in only a few dozen or even fewer reliable labeled samples in most areas. Under this "small sample" condition, traditional machine learning models (such as random forests and support vector machines) and even data-driven deep learning models are prone to overfitting. The models tend to memorize noise from a limited number of samples rather than learn generalization rules, leading to a significant reduction in the reliability and robustness of their classification results, making it difficult to support management decisions. Secondly, the ambiguity of mixed and transitional categories places higher demands on feature identification. "Naturally expanded areas after human intervention," as typical ecological continuums, occupy a transitional zone between planted forests and purely natural forests in the feature space. Existing classifiers are mostly "hard classifiers," requiring each pixel to be forcibly assigned to a discrete category, making it difficult to properly characterize the fuzzy attributes of these categories, leading to a large number of misclassified transitional pixels. Furthermore, the disconnect between technical processes and domain management rules restricts operational applications. Most existing methods remain at the "pixel-feature" calculation level, failing to incorporate operational standards for mangrove restoration from management departments (such as minimum continuous area, boundary regularity, etc.). This not only results in inefficient classification results but also introduces subjective bias, failing to meet the efficiency requirements of routine, operational monitoring.

[0061] To address the aforementioned challenges, various solutions have been developed in the existing technology field. Among them, the one most similar to this embodiment is "mangrove afforestation site mapping based on multi-temporal Sentinel-2 imagery and random forest." This solution extracts multi-seasonal spectral and textural features and uses random forest for classification, showing some effectiveness in ideal scenarios with sufficient data. However, its inherent limitations are significant in complex real-world scenarios: First, the performance of random forest degrades sharply with small samples because its bootstrapping-based ensemble mechanism is extremely unstable when data is scarce, making it unreliable in areas with only a few known plantation patches (e.g., dozens); second, its hard classification logic struggles to identify the transitional category of "natural expansion after human intervention"; and third, the entire process lacks an automated module for integration with management rules such as area thresholds, resulting in insufficient practicality. These shortcomings collectively limit the potential for operational portability and large-scale application of this solution.

[0062] In summary, there is a clear technological gap in the current field: there is an urgent need for an end-to-end technical solution that can work stably under conditions with very few known samples, accurately identify mixed artificial and natural features, automatically execute embedded domain management rules, and intelligently integrate multi-source heterogeneous features.

[0063] Despite advancements in existing technologies, the following key limitations remain in complex coastal wetland environments:

[0064] 1. Strong sample dependence: Existing supervised learning methods experience a sharp decline in performance under small sample conditions. The bootstrap sampling mechanism of random forests cannot guarantee the diversity of base learners when samples are scarce, resulting in a significant reduction in the model's generalization ability.

[0065] 2. Rigid category recognition: Traditional hard classifiers cannot effectively recognize ecological continuums such as "natural expansion areas after human intervention". The mandatory discrete classification boundary is inherently contradictory to the ecological gradual change process, resulting in the lack of transition type recognition.

[0066] 3. Disconnection between management rules: The existing technical process lacks an automatic docking mechanism with the management rules for mangrove restoration (such as the minimum continuous area threshold), which is inefficient and highly subjective.

[0067] The purpose of this application is to provide an innovative solution to address the core shortcomings of existing mangrove growth type identification technologies, such as poor generalization ability under small sample conditions, difficulty in effectively identifying ecological transition types, inability to connect with management rules, and inefficiency in multi-source feature fusion. By constructing an intelligent classification system based on an advanced machine learning framework, this application aims to achieve the following: First, maintaining high robustness and high accuracy in classification even with a very small number of known plantation samples; second, accurately identifying and quantifying the ecological continuum of "natural expansion area after human intervention," truly reflecting the gradual process of mangrove restoration; and third, embedding management rules such as minimum area thresholds into the automated processing flow, connecting the key links from remote sensing image analysis to management decision-making applications, ultimately forming a scientific, reliable, efficient, and practical technical system that provides directly usable decision support tools for mangrove restoration effectiveness assessment, accurate blue carbon accounting, and refined ecological restoration management.

[0068] Reference Figure 1 This application provides a method for identifying naturally propagated and artificially planted mangroves, specifically including the following steps S100~S160:

[0069] S100: Extract the comprehensive feature matrix of the study area within the target year;

[0070] S110: Identify multiple initial sampling points within known plantation patches in the study area;

[0071] S120: Extract the initial feature matrix corresponding to each initial sampling point from the comprehensive feature matrix;

[0072] S130: Based on the initial feature matrix, determine multiple target sampling points within the newly added mangrove area in the study area;

[0073] S140: Extract the target feature matrix corresponding to each target sampling point from the comprehensive feature matrix;

[0074] S150: Input each of the target feature matrices into the pre-trained ensemble learning model to obtain the comprehensive probability distribution output by the pre-trained ensemble learning model;

[0075] S160: Based on the comprehensive probability distribution, the newly added mangrove area is classified into at least one of the following: natural propagation, artificial planting, or natural expansion after human intervention.

[0076] Optionally, the extraction of the comprehensive feature matrix of the study area within the target year includes the following steps:

[0077] The spectral features of the study area in the target year, the structural features measured by synthetic aperture radar, the topographic background features, the temporal harmonic features, and the multi-scale texture features are extracted as the comprehensive feature matrix.

[0078] Optionally, determining multiple target sampling points within the newly added mangrove area in the study area based on the initial feature matrix includes the following steps:

[0079] For each unlabeled pixel in the newly added mangrove area, the cosine similarity between the corresponding feature vector and the initial feature matrix of each initial sampling point is calculated, and the maximum value is taken; wherein, the comprehensive feature matrix includes the feature vector of the unlabeled pixel;

[0080] If the cosine similarity of the maximum value is greater than the first set threshold, then the corresponding unlabeled pixel is determined as the target sampling point for the expanded positive sample;

[0081] If the cosine similarity of the maximum value is less than the second set threshold, then the corresponding unlabeled pixel is determined as the target sampling point of the confidence negative sample; wherein, the second set threshold is less than the first set threshold.

[0082] Optionally, the ensemble learning model includes an XGBoost classifier and a random forest classifier;

[0083] The pre-trained ensemble learning model is obtained through the following steps:

[0084] The XGBoost classifier and the random forest classifier are trained in parallel; both classifiers use five-fold cross-validation to optimize hyperparameters.

[0085] Based on the weighted F1 scores on the validation set, the optimal ensemble weights for the XGBoost classifier and the random forest classifier are dynamically calculated respectively.

[0086] The probability distributions output by the XGBoost classifier and the random forest classifier are weighted and summed according to their respective optimal ensemble weights to obtain the comprehensive probability distribution; wherein, the probability distribution represents the probability that each unlabeled pixel belongs to a human-intervention patch.

[0087] Optionally, the comprehensive probability distribution is calculated using the following expression:

[0088] ;

[0089] in, For the aforementioned comprehensive probability distribution, The optimal ensemble weights for the XGBoost classifier are... The probability distribution output by the XGBoost classifier is... The optimal ensemble weights for the random forest classifier are... This represents the probability distribution output by the random forest classifier.

[0090] Optionally, classifying the newly added mangrove area into at least one of natural propagation, artificial planting, or natural expansion after human intervention based on the comprehensive probability distribution includes the following steps:

[0091] The pixel coordinates classified as human intervention are selected based on the comprehensive probability distribution.

[0092] Several candidate artificial forest patches are determined based on the distance between the coordinates of the pixels involved in each human intervention.

[0093] If the continuous area of ​​the candidate plantation patch reaches a preset area threshold, it is determined to be an artificially planted patch.

[0094] If the continuous area of ​​the candidate plantation patch is less than the preset area threshold, then the minimum Euclidean distance to the centroid of each plantation patch is calculated.

[0095] If the minimum Euclidean distance is less than or equal to a preset spatial proximity threshold, then the corresponding candidate plantation patch is determined to be a patch that has naturally expanded after human intervention.

[0096] If the minimum Euclidean distance is greater than the preset spatial proximity threshold, then the corresponding candidate plantation patch is determined to be a naturally proliferating patch.

[0097] Optionally, determining a plurality of candidate artificial forest patches based on the distance between the coordinates of each of the human-interventioned pixels includes the following steps:

[0098] The DBSCAN algorithm is used to cluster the coordinates of each human intervention cell and remove noise coordinates to obtain several non-noise clusters. Each non-noise cluster is a candidate artificial forest patch.

[0099] The following section will provide a detailed introduction and explanation of the solutions in the embodiments of this application, using specific application examples.

[0100] This embodiment provides a method and system for intelligent identification of newly added mangrove growth types based on multi-source remote sensing and ensemble learning. The core lies in constructing an end-to-end technical framework that includes five-dimensional feature matrix extraction, two-layer ensemble learning modeling, and deep fusion of management rules, thus overcoming the application bottlenecks of existing methods in scenarios such as small sample sizes, complex categories, and management disconnect. Figure 2 As shown, the complete technical solution of this embodiment follows a progressive logic of "data-driven, model-intelligent, and rule-validated".

[0101] Specifically, this embodiment includes the following technical solutions:

[0102] Step S1: New region extraction and anchor point sample preparation.

[0103] High-quality initial learning sample preparation. Based on real new additions that exist in the target year but not in the baseline year, import vector files of known artificial afforestation patches that have been verified in the field or confirmed by high-resolution imagery. For each patch, the system generates a regular sampling grid with 10-meter intervals within its geometric boundaries as "seeds" for subsequent small-sample learning.

[0104] Step S2: Automated extraction of the five-dimensional comprehensive feature matrix.

[0105] Annual spectral characteristics are based on Sentinel-2 L2A level surface reflectance products. Median composite images of the growing season (May-October) are generated for each target year, effectively suppressing the influence of cloud cover and transient anomalies. Reflectance values ​​for 10 bands (including the vegetation-sensitive red-edge bands B5, B6, B7, and B8A) and four key vegetation indices are extracted from the composite images: Normalized Differential Vegetation Index (NDVI) reflects overall greenness; Enhanced Vegetation Index (EVI) is insensitive to saturation effects in high biomass areas; and Bare Soil Index (BSI) and Modified Normalized Difference Water Index (MNDWI) help distinguish between bare land and tidal gullies, respectively.

[0106] SAR structural features were derived using C-band synthetic aperture radar data from Sentinel-1. Preprocessing included precise orbit correction, thermal noise removal, topographic radiometric correction, and Refined Lee filtering to suppress speckle noise. The system generated quarterly synthetic images (January–March, April–June, July–September, and October–December), extracting the backscattering coefficients (in dB) of VV and VH polarizations. VV polarization is sensitive to vegetation vertical structure, while VH polarization better reflects vegetation volume scattering and water content information. The average values ​​for the four quarters were used to provide structural information complementary to the optical data.

[0107] The topographic background features are based on 30-meter resolution Space Shuttle Radar Topography Mission (SRTM) digital elevation data. This includes elevation values, slope (degrees), aspect (converted to sine and cosine values ​​to eliminate circumferential discontinuities), and topographic humidity index. These reflect the site conditions for mangrove growth; intertidal elevation and micro-topography directly influence the natural dispersal process of mangroves and the selection of artificial planting sites.

[0108] High-frequency time-series harmonic features were constructed based on Sentinel-2 images of all cloud cover below 30% during the study period, with a complete NDVI time series built for each sampling point. Harmonic analysis was used to fit a model containing three harmonic components, from which eight key time-series features were extracted: the maximum and minimum values ​​of the NDVI sequence, interannual amplitude (reflecting growth vigor), first harmonic phase (indicating the time of peak occurrence during the growing season), first harmonic amplitude (characterizing the intensity of seasonal variation), root mean square error of fitting (reflecting the noise level of the sequence), NDVI coefficient of variation, and the mangrove variation index newly defined in this embodiment.

[0109] ;

[0110] Where A1 is the first harmonic amplitude, and SC is the coefficient of variation (standard deviation divided by mean) among the three harmonic amplitudes. This index comprehensively reflects the seasonal regularity and interannual stability of vegetation growth: artificially managed mangroves usually exhibit high and stable MVI values; while naturally growing mangroves are more affected by climate and tidal fluctuations, and their MVI values ​​are relatively low and fluctuate more.

[0111] Multi-scale texture features were extracted using the PCA-GLCM method. First, the annual composite images of the Sentinel-2 near-infrared band (B8) and the Sentinel-1 VV and VH bands were converted into 8-bit grayscale images. For each grayscale image, a sliding window was used to calculate the gray-level co-occurrence matrix, extracting four basic texture measures: contrast, correlation, energy, and homogeneity. GLCM features at three scales (3×3, 5×5, and 7×7) were calculated. Then, principal component analysis was performed on multiple texture measures, retaining the first principal component as the "comprehensive texture value" for that year and data source to achieve efficient compression and redundancy removal of texture information.

[0112] Step S3: Intelligent sample augmentation and ensemble learning model construction based on anchor points.

[0113] The core idea of ​​intelligent sample augmentation is to utilize the prior geometric structure of the feature space. First, a feature subset F_anchor of anchor point samples is separated from the comprehensive feature matrix. For each unlabeled pixel in the newly added region, the cosine similarity between its feature vector and the feature vectors of all anchor points is calculated, and the maximum value S_max is taken. The formula for calculating cosine similarity is:

[0114] ;

[0115] Here, x is the feature vector of an unlabeled pixel, and y is the feature vector of an anchor sample, ranging from -1 to 1, with larger values ​​indicating greater feature similarity. Pixels with S_max > θ are automatically labeled as "expanded positive samples," considered homogeneous with known plantation heights. Simultaneously, the top 10% of pixels with the smallest S_max are selected as high-confidence negative samples (pure natural growth candidates). This process significantly expands the size of the effective training samples without increasing the cost of manual annotation.

[0116] The ensemble learning model employs an innovative two-layer weighted architecture, fully leveraging the precise gradient optimization capabilities of XGBoost and the robust noise resistance of Random Forest. The first layer trains two strong heterogeneous base classifiers in parallel: the XGBoost classifier is set with a learning rate of 0.1, a maximum tree depth of 6, and 100 trees, and early stopping is enabled to prevent overfitting; the Random Forest classifier has 200 trees, and the maximum number of features is the square root of the total number of features. Both classifiers use five-fold cross-validation to optimize hyperparameters. The second layer implements a meta-learning strategy: based on the weighted F1 scores on the validation set, the optimal ensemble weights of the two base classifiers are dynamically calculated. The final predicted probability of the ensemble learning model is generated by a weighted average.

[0117] ;

[0118] The weights sum to 1. After model training, it is applied to the feature matrix of all sampling points within the newly added mangrove area to obtain the probability distribution of each pixel belonging to direct human intervention. A confidence-based decision rule is adopted: if the maximum probability value exceeds the threshold of 0.7, the corresponding category label is assigned; otherwise, it is marked as "uncertain". This soft classification output preserves important uncertainty information for subsequent spatial analysis and rule verification.

[0119] Step S4: Manage the embedded category assistance within the rules.

[0120] The first step in spatial processing is density-based clustering analysis. The coordinates of all pixels initially classified as "direct human intervention" are extracted, and the DBSCAN algorithm is applied for spatial clustering. DBSCAN parameters are set to a neighborhood radius eps = 20 meters (approximately 2 Sentinel-2 pixel distances) and a minimum number of points minPts = 5. This algorithm can discover density-connected regions of arbitrary shapes and identify sparse discrete points as noise. Each non-noise cluster forms a candidate artificial forest patch, its spatial extent defined by the convex hull or minimum bounding rectangle of all pixels within the cluster.

[0121] The core of patch compliance determination is area threshold verification. For each candidate plantation patch, the number of effective pixels it contains is counted, multiplied by the actual area of ​​the pixels (e.g., 10 meters of resolution corresponds to 100 square meters) to obtain the continuous area of ​​the patch. Based on relevant industry standards, a minimum area threshold A_min = 400 square meters is set. Patches meeting the area standard (Area ≥ A_min) are officially recognized as "plantation patches". For small patches with insufficient area, the minimum Euclidean distance D_min between the centroid of the small patch and the centroids of all qualified large patches is calculated. A spatial proximity threshold D_thresh = 50 meters is set, based on ecological principles: the effective dispersal distance of mangrove seeds is usually limited. If D_min ≤ D_thresh, the small patch is considered to have formed naturally from the seed dispersal of neighboring large patches and is reclassified as "natural expansion after human intervention"; if D_min > D_thresh, it is considered an isolated misjudgment or the smallest planting unit that does not meet management requirements and is reclassified as "natural". This process takes into full account the ecological process and spatial continuity of mangrove restoration.

[0122] In summary, this embodiment includes the following key technical solutions:

[0123] 1. A sample intelligent augmentation method based on anchor point guidance and feature space similarity analysis.

[0124] In mangrove growth type identification, addressing the bottleneck of scarce known plantation samples, this embodiment innovatively proposes an intelligent sample augmentation method. This method first uses a small number of confirmed plantation patches as "anchor points," extracting their multi-dimensional features to construct a reference feature space. Then, it calculates the similarity metrics (such as cosine similarity and Mahalanobis distance) between all unlabeled pixels in the newly added area and these anchor points in the feature space. Finally, pixels with similarity exceeding a preset threshold (e.g., 0.85) are automatically labeled as "expanded positive samples," thereby constructing a training sample set with significantly improved scale and quality. The core of this method lies in utilizing the prior geometric structure of the feature space, rather than relying on extensive manual annotation, thus providing a reliable data foundation for subsequent machine learning models under small sample conditions.

[0125] 2. An adaptive harmonic feature extraction method and a new index (MVI) for mangrove phenological patterns.

[0126] To robustly extract unique growth rhythm information of mangroves from incomplete satellite time-series data affected by cloud cover and tides, this embodiment proposes an improved harmonic analysis method. This method not only fits multiple harmonic components of the NDVI time series and extracts conventional indicators such as amplitude, phase, and fitting error, but more importantly, it innovatively defines and calculates a new time-series characteristic index, namely the mangrove variation index, which characterizes the seasonal stability and regularity of vegetation growth.

[0127] 3. A two-layer weighted ensemble learning classification framework that integrates XGBoost and random forest.

[0128] To address overfitting and model instability issues in scenarios with small sample sizes, high dimensionality, and complex categories, this embodiment proposes a novel ensemble learning classification framework. This framework is not a simple model averaging but employs a two-layer weighted fusion strategy: The first layer trains two strong heterogeneous base classifiers, XGBoost and Random Forest, in parallel; the second layer uses a meta-learning approach to dynamically allocate the weights of the two base classifiers in the final decision based on validation set performance (e.g., assigning higher weights to XGBoost to leverage its accurate gradient optimization capabilities while retaining Random Forest for enhanced robustness), and then performs a weighted average of their class probability outputs to obtain the final soft classification result.

[0129] 4. A method for automatic verification of spatial post-processing and area thresholds for embedded management rules.

[0130] To achieve the direct transformation of technological achievements into management operations, this embodiment deeply embeds domain management rules (minimum continuous area requirements) into the classification processing flow. The method first uses a spatial clustering algorithm (such as DBSCAN) to aggregate the initially classified "direct human intervention" pixels into candidate plantation patches; then, it automatically calculates the continuous area of ​​each patch; finally, it automatically determines the validity of the patch based on a preset area threshold (such as 400 square meters): those that meet the threshold are marked as "qualified plantation forests," those that do not meet the threshold but are adjacent to large qualified patches are reclassified as "natural expansion after human intervention," and isolated small patches are marked as "natural."

[0131] 5. PCA-GLCM texture fusion and five-dimensional collaborative feature engineering system with multi-source heterogeneous features.

[0132] To overcome the drawbacks of simply stacking multi-source features, this embodiment constructs a systematic five-dimensional collaborative feature engineering system. Its core innovation lies in the processing of texture features: for each year's optical and SAR grayscale imagery, a multi-scale gray-level co-occurrence matrix (GLCM) is calculated to obtain multiple texture measures. Principal component analysis (PCA) is then performed on these multiple texture measures, retaining the first principal component as the comprehensive texture feature of the data for that year. This method achieves efficient compression and redundancy removal of texture information. This feature system ultimately synergistically integrates five dimensions: annual spectral features, SAR structural features, topographic background features, high-frequency temporal harmonic features, and PCA-GLCM texture features.

[0133] Beneficial effects:

[0134] In terms of technical performance, this approach overcomes the small sample constraint. Existing random forest methods suffer from drastic performance degradation when samples are scarce. However, this embodiment maximizes the utilization and effective generalization of discriminative information from limited known samples through anchor-guided intelligent sample augmentation and the XGBoost-random forest two-layer weighted ensemble framework. This enables stable and reliable high-precision classification even with only a few dozen known plantation patches, addressing the fundamental limitation of existing technologies in applying to regions with scarce samples.

[0135] In scientific understanding and management applications, existing technologies can only perform a rigid dichotomy of "artificial / natural," while this embodiment deeply internalizes the management department's standards for determining the area of ​​plantations into the core constraints and spatial aggregation criteria of the classification model. Through soft classification output and a dual criterion of "feature + space," it achieves effective identification and quantification of the key ecological transition type of "natural expansion after human intervention."

[0136] Reference Figure 3 This application provides a system for identifying naturally propagated and artificially planted mangroves, including:

[0137] The comprehensive feature extraction unit is used to extract the comprehensive feature matrix of the study area within the target year;

[0138] The first anchor point unit is used to determine multiple initial sampling points within known plantation patches in the study area;

[0139] An initial feature extraction unit is used to extract the initial feature matrix corresponding to each initial sampling point from the comprehensive feature matrix;

[0140] The second anchor point unit is used to determine multiple target sampling points within the newly added mangrove area in the study area based on the initial feature matrix;

[0141] The target feature extraction unit is used to extract the target feature matrix corresponding to each target sampling point from the comprehensive feature matrix;

[0142] The probability prediction unit is used to input each of the target feature matrices into the pre-trained ensemble learning model to obtain the comprehensive probability distribution output by the pre-trained ensemble learning model.

[0143] The regional division unit is used to divide the newly added mangrove area into at least one of the following categories based on the comprehensive probability distribution: natural propagation, artificial planting, or natural expansion after human intervention.

[0144] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0145] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0146] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0147] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0148] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0149] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0150] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0151] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0152] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0153] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for identifying naturally propagated and artificially planted mangroves, characterized in that, The method includes the following steps: Extract the comprehensive feature matrix of the study area within the target year; Multiple initial sampling points were identified within known plantation patches in the study area; Extract the initial feature matrix corresponding to each initial sampling point from the comprehensive feature matrix; Based on the initial feature matrix, multiple target sampling points are determined within the newly added mangrove area in the study area; Extract the target feature matrix corresponding to each target sampling point from the comprehensive feature matrix; Each of the target feature matrices is input into a pre-trained ensemble learning model to obtain the comprehensive probability distribution output by the pre-trained ensemble learning model. Based on the comprehensive probability distribution, the newly added mangrove area is classified into at least one of the following: natural propagation, artificial planting, or natural expansion after human intervention.

2. The method for identifying naturally propagated and artificially planted mangroves according to claim 1, characterized in that, The extraction of the comprehensive feature matrix of the study area within the target year includes the following steps: The spectral features of the study area in the target year, the structural features measured by synthetic aperture radar, the topographic background features, the temporal harmonic features, and the multi-scale texture features are extracted as the comprehensive feature matrix.

3. The method for identifying naturally propagated and artificially planted mangroves according to claim 1, characterized in that, The process of determining multiple target sampling points within the newly added mangrove area in the study area based on the initial feature matrix includes the following steps: For each unlabeled pixel in the newly added mangrove area, the cosine similarity between the corresponding feature vector and the initial feature matrix of each initial sampling point is calculated, and the maximum value is taken; wherein, the comprehensive feature matrix includes the feature vector of the unlabeled pixel; If the cosine similarity of the maximum value is greater than the first set threshold, then the corresponding unlabeled pixel is determined as the target sampling point for the expanded positive sample; If the cosine similarity of the maximum value is less than the second set threshold, then the corresponding unlabeled pixel is determined as the target sampling point of the confidence negative sample; wherein, the second set threshold is less than the first set threshold.

4. The method for identifying naturally propagated and artificially planted mangroves according to claim 3, characterized in that, Ensemble learning models include XGBoost classifiers and random forest classifiers; The pre-trained ensemble learning model is obtained through the following steps: The XGBoost classifier and the random forest classifier are trained in parallel; both classifiers use five-fold cross-validation to optimize hyperparameters. Based on the weighted F1 scores on the validation set, the optimal ensemble weights for the XGBoost classifier and the random forest classifier are dynamically calculated respectively. The probability distributions output by the XGBoost classifier and the random forest classifier are weighted and summed according to their respective optimal ensemble weights to obtain the comprehensive probability distribution; wherein, the probability distribution represents the probability that each unlabeled pixel belongs to a human-intervention patch.

5. The method for identifying naturally propagated and artificially planted mangroves according to claim 4, characterized in that, The comprehensive probability distribution is calculated using the following expression: ； in, For the aforementioned comprehensive probability distribution, The optimal ensemble weights for the XGBoost classifier are... The probability distribution output by the XGBoost classifier is... The optimal ensemble weights for the random forest classifier are... This represents the probability distribution output by the random forest classifier.

6. A method for identifying naturally propagated and artificially planted mangroves according to any one of claims 1 to 5, characterized in that, The step of classifying the newly added mangrove area into at least one of natural propagation, artificial planting, or natural expansion after human intervention based on the comprehensive probability distribution includes the following steps: The pixel coordinates classified as human intervention are selected based on the comprehensive probability distribution. Several candidate artificial forest patches are determined based on the distance between the coordinates of the pixels involved in each human intervention. If the continuous area of ​​the candidate plantation patch reaches a preset area threshold, it is determined to be an artificially planted patch. If the continuous area of ​​the candidate plantation patch is less than the preset area threshold, then the minimum Euclidean distance to the centroid of each plantation patch is calculated. If the minimum Euclidean distance is less than or equal to a preset spatial proximity threshold, then the corresponding candidate plantation patch is determined to be a patch that has naturally expanded after human intervention. If the minimum Euclidean distance is greater than the preset spatial proximity threshold, then the corresponding candidate plantation patch is determined to be a naturally proliferating patch.

7. The method for identifying naturally propagated and artificially planted mangroves according to claim 6, characterized in that, The process of determining several candidate artificial forest patches based on the distance between the coordinates of each human-involved pixel includes the following steps: The DBSCAN algorithm is used to cluster the coordinates of each human intervention cell and remove noise coordinates to obtain several non-noise clusters. Each non-noise cluster is a candidate artificial forest patch.

8. A system for identifying naturally propagated and artificially planted mangroves, characterized in that, The system includes: The comprehensive feature extraction unit is used to extract the comprehensive feature matrix of the study area within the target year; The first anchor point unit is used to determine multiple initial sampling points within known plantation patches in the study area; An initial feature extraction unit is used to extract the initial feature matrix corresponding to each initial sampling point from the comprehensive feature matrix; The second anchor point unit is used to determine multiple target sampling points within the newly added mangrove area in the study area based on the initial feature matrix; The target feature extraction unit is used to extract the target feature matrix corresponding to each target sampling point from the comprehensive feature matrix; The probability prediction unit is used to input each of the target feature matrices into the pre-trained ensemble learning model to obtain the comprehensive probability distribution output by the pre-trained ensemble learning model. The regional division unit is used to divide the newly added mangrove area into at least one of the following categories based on the comprehensive probability distribution: natural propagation, artificial planting, or natural expansion after human intervention.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 7.

Images