A hierarchical modular ecosystem mapping framework for high intensity human activity areas

Abstract

The application discloses a kind of layered modular ecological system mapping framework construction methods for high-intensity human activity area, including constructing training sample library suitable for large-scale ecological system mapping, constructing hierarchical classification structure in line with ecological system type cognitive rule, constructing space-time characteristics, outputting initial ecological system spatial distribution map, refining and correcting initial ecological system spatial distribution map, obtaining the final ecological system map for high-intensity human activity area five steps.The method can improve the class separability of complex regional ecological system mapping, reduce model classification uncertainty, and enhance the ability to distinguish similar subcategories of ecological properties, accurately identify and depict the spatial distribution pattern of various ecological systems.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing information processing and ecosystem mapping technology, and particularly relates to a method for constructing a hierarchical modular ecosystem mapping framework for areas with high human activity. Background Technology

[0002] High-intensity human activity exerts continuous and significant pressure on ecosystems, leading to the degradation of ecosystem structure and function. Therefore, accurately identifying and characterizing the spatial distribution patterns of various ecosystems is fundamental to assessing regional ecological integrity and formulating scientific protection strategies. In recent years, the rapid development of remote sensing technology has provided an economical, efficient, and repeatable technical means for large-scale ecosystem monitoring. Existing methods typically rely on Earth observation data, combined with supervised classification methods such as random forests, support vector machines, and gradient boosting trees, to identify ecosystem types such as forests, cultivated land, water bodies, wetlands, and built-up land. With the rise of cloud computing platforms such as Google Earth Engine, large-scale ecosystem mapping integrating multi-source time-series remote sensing data has become the mainstream method. However, under the combined influence of rapid urbanization and land use transformation, regional ecosystems generally exhibit strong fragmentation characteristics and significant spatiotemporal heterogeneity. Although existing land cover or ecosystem products can reflect the overall regional pattern relatively well, they often still fall short of meeting the practical needs of refined management and ecological assessment in terms of thematic classification accuracy, ecological subclass identification capability, and spatial detail representation. Therefore, in order to improve the class separability of ecosystem mapping in complex regions, reduce the uncertainty of model classification, and enhance the ability to distinguish subclasses with similar ecological attributes (such as evergreen forests and deciduous forests, rivers and lakes, inland wetlands and coastal wetlands), it is urgent to construct an ecosystem mapping framework with good reusability and portability.

[0003] From the perspective of the technical implementation mechanism of classification strategies, ecosystem mapping mainly includes Flat Classification Strategy (FCS) and Hierarchical Classification Strategy (HCS). FCS has a relatively intuitive and simple process, typically inputting all target ecosystem categories into a single classifier at once. This strategy learns the decision boundaries between various ecosystems directly in a multi-dimensional feature space through the model, thereby achieving the identification and classification of all types at once. In contrast, HCS is based on a comprehensive guidance of prior information, phenological characteristics, spatial context, and expert knowledge. It reconstructs the complex global classification task into a logically rigorous hierarchical structure and breaks it down into a series of sequentially executed, relatively independent, continuous subtasks. Its core idea is to first complete the discrimination of higher-level major categories based on the ecological differences, spectral response patterns, and hierarchical relationships between categories, and then gradually refine them into more specific subcategories, thus forming a progressive classification process from coarse to fine, layer by layer. Through this organizational approach, HCS not only intuitively reflects the inherent hierarchy of ecosystem classification systems and enhances the consistency between the classification process and ecological understanding, but also effectively distinguishes land cover categories that are distinctly different in ecological function and thematic definition, yet exhibit high spectral similarity in remote sensing imagery. Especially in scenarios with numerous ecosystem types and complex inter-category relationships, the hierarchical classification strategy provides an extremely clear classification path and organizational framework for subsequent refined identification and mapping.

[0004] Although both FCS and HCS are widely used in ecosystem mapping, they still have certain technical limitations in practice. FCS places all target categories in the same feature space for unified discrimination, making its methodology relatively straightforward. However, as the number of target categories increases, the overlap of different ecosystems in phenology, geography, and spectral characteristics increases significantly, leading to a decline in classification performance. When facing complex regions with strong spatial heterogeneity and subtle spectral differences, the classifier faces a significantly increased difficulty in distinguishing all categories simultaneously, making class confusion highly likely. In contrast, while HCS reduces the complexity of a single classification task through hierarchical organization, its framework is highly dependent on prior knowledge, expert experience, and the rational design of hierarchical rules. More importantly, the overall performance of HCS is extremely sensitive to the accuracy of higher-level classifications. Once errors or misclassifications are introduced in the early stages of upper-level classification, they will constrain lower-level refinement, leading to irreversible downward error propagation. These errors accumulate in subsequent refinement processes and are difficult to correct. Furthermore, the effectiveness of HCS classification is significantly reduced when dealing with land cover types that have the exact same spectral characteristics but perform different ecological functions.

[0005] In summary, existing methods for constructing hierarchical modular ecosystem mapping frameworks for areas with high-intensity human activity have the following problems: (1) Categories with similar spectral characteristics but different ecological functions are difficult to distinguish effectively; (2) The prediction performance of a single classifier is unstable when dealing with heterogeneous landscapes; (3) In the traditional hierarchical classification architecture, inaccurate classification at the upper level can easily lead to irreversible downward error propagation; (4) Complex wetland and forest subclasses are difficult to identify and extract in detail. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention proposes a hierarchical and modular ecosystem mapping framework construction method. This method deeply integrates hierarchical classification strategies, probabilistic ensemble methods, and knowledge rule-based pixel and object optimization algorithms, which can effectively alleviate the above-mentioned technical bottlenecks and significantly improve the accuracy, robustness, and cross-regional transferability of detailed ecosystem mapping in complex areas.

[0007] The technical problem solved by this invention is achieved through the following technical solution: A method for constructing a hierarchical modular ecosystem mapping framework for areas with high human activity includes the following steps: Step S1: Construct a training sample library suitable for large-scale ecosystem mapping by combining automatically generated training samples with visual interpretation samples; Step S2: Construct a hierarchical classification structure that conforms to the cognitive rules of ecosystem types, and decompose the current ecosystem mapping task into multiple interconnected sub-tasks. Step S3: Acquire multi-source remote sensing data and construct spatiotemporal features; Step S4: By integrating hierarchical classification strategies and probabilistic ensemble methods, the initial ecosystem spatial distribution map is generated by combining the outputs of multiple classification models at each level node; Step S5: By fusing pixel-scale features, object-scale features, and knowledge rule optimization methods, the initial ecosystem spatial distribution map is refined and corrected to obtain the final ecosystem map oriented towards areas with high-intensity human activity.

[0008] Further, step S1 includes the following steps: Step S1.1: Input multi-source reference data for mapping ecosystems in areas with high human activity, unify the spatial resolution and coordinate system of the multi-source reference data, and then identify candidate sample regions based on stable pixels with consistent classification in each reference year, cross-consistency of multiple products, and neighborhood filtering results, and automatically generate training samples based on the candidate sample regions. Step S1.2: For regions in the candidate sample area that do not meet the conditions for automatic sample generation, perform visual interpretation based on remote sensing images and their temporal index features to form target interpretation samples; Step S1.3: Based on automatically generated samples and target interpretation samples, construct a training sample library suitable for large-scale ecosystem mapping.

[0009] Furthermore, in step S1.1, the multi-source reference data for mapping ecosystems in areas with high-intensity human activity is multi-source land cover or land use product data; in step S1.2, the remote sensing images used for visual interpretation can be Landsat images, Sentinel-2 images, high-resolution remote sensing images, or Gaofen series satellite images, and the temporal index features are NDVI temporal features and NDWI temporal features obtained after temporal smoothing.

[0010] Furthermore, in step S2, a hierarchical classification structure conforming to the cognitive laws of ecosystem types includes multiple levels L1~Li, i=2,3,4,... Starting from level L1, all ecosystem types in the target area are initially classified, and then level L2 classification is carried out based on the classification results of level L1. Each level is then classified in the same way, thus finally generating a high-precision initial ecosystem mapping result.

[0011] Furthermore, in step S2, the hierarchical classification structure conforming to the cognitive patterns of ecosystem types includes five levels, namely L1, L2, L3, L4, and L5. Level L1 includes permanent water, permanent dry water, and periodic water categories; Level L2 includes vegetation cover and no vegetation cover categories refined from the permanent dry water and periodic water categories in Level L1; Level L3 includes woody vegetation and herbaceous vegetation categories refined from the vegetation cover category in Level L2 corresponding to the permanent dry water category in Level L1; Level L4 includes the water body category corresponding to the permanent water category in Level L1, and the woody vegetation category refined from the woody vegetation category in Level L3, including forest and shrubland categories. Level L3 includes grassland and dryland categories refined from herbaceous vegetation categories; built-up land and bare land categories refined from the no-vegetation-cover category corresponding to the permanent waterless category in Level L2; paddy fields and marshes categories refined from the vegetation-cover category corresponding to the periodic water category in Level L2; and flooded ponds categories generated by the no-vegetation-cover category corresponding to the periodic water category in Level L2. Level L5 includes rivers, lakes, and ocean categories refined from the water body category in Level L4; evergreen forest and deciduous forest categories refined from the forest category; inland marshes and coastal marshes categories refined from the marshland category; and tidal flats and mudflats categories refined from the flooded pond category.

[0012] Further, step S3 specifically involves acquiring multi-source optical remote sensing data, Sentinel-1 dual-polarization data, and digital elevation model data for the required year; preprocessing the multi-source optical remote sensing data to remove invalid observations affected by clouds, cloud shadows, and snow / ice; constructing a time series using monthly median synthesis; reconstructing missing months using linear interpolation; extracting spectral index features, maximum value features, temporal features, and texture features based on the reconstructed optical time series; extracting polarization features based on Sentinel-1 dual-polarization data; and extracting terrain features based on digital elevation model data. The multi-source optical remote sensing data includes HLSL30 data and HLSS30 data. The spectral index features include one or more of NDVI, EVI, SAVI, NDWI, MNDWI, LSWI, NBR, and NBR2 calculated based on the reconstructed optical time series. Missing months are reconstructed using linear interpolation.

[0013] Further, step S4 includes the following steps: Step S4.1: Before performing classification, the spatiotemporal features constructed in step S3 are uniformly normalized to the range of [0,1]. Then, a machine learning base model with complementary features in feature discrimination is constructed and configured. Step S4.2: Based on the trained machine learning base model, generate the category probability distribution map of each pixel in the image to be classified, and then calculate the probability value of the same pixel belonging to each category under each model. At any level of classification node, assign the category with the highest probability value to the corresponding pixel as the final prediction result of that node. The entire classification process strictly follows the hierarchical structure classification. Through step-by-step subdivision, the initial ecosystem spatial distribution map is finally generated.

[0014] Further, step S5 includes the following steps: Step S5.1: Based on the seasonal NDVI difference, a preset threshold is used to distinguish between evergreen forests and deciduous forests; Step S5.2: Based on the coastline data, construct an inland buffer zone and distinguish between coastal wetlands / coastal water bodies and inland wetlands / inland water bodies; Step S5.3: Convert inland water bodies into vector objects, extract geometric features such as compactness, density, and rectangle fit at the object level, and achieve river-lake separation based on combination rules.

[0015] Furthermore, in step S4.1, the machine learning base models are the Support Vector Machine (SVM) model, the Random Forest (RF) model which balances model stability and generalization performance, and the Gradient Boosting Tree (GPB) model which achieves a better balance between classification performance and computational efficiency. The GPB model can be GTB, XGBoost, or LightGBM.

[0016] Furthermore, in step S4.2, when calculating the probability value of the same pixel belonging to each category under each model, the probability value can be the equally weighted average probability value, or a weighted fusion probability value based on the cross-validation accuracy can be set for different machine learning base models.

[0017] Compared with existing technologies, the beneficial technical effects of the present invention are as follows: By employing a hierarchical classification strategy combined with a probabilistic ensemble method for joint modeling, the overall accuracy and stability of ecosystem classification in complex regions were significantly improved. Compared with single-model methods, the overall accuracy was increased by 2.50–5.41 percentage points, and the Kappa coefficient was improved by 0.0301–0.0654.

[0018] By introducing refined rules based on phenological, morphological, and locational knowledge, the ability to distinguish between subcategories such as rivers / lakes, evergreen / deciduous forests, and inland / coastal wetlands has been enhanced, making up for the shortcomings of pure spectral classification in expressing differences in ecological functions.

[0019] (3) For the major ecosystems involved in the industry, an overall accuracy of over 90% was achieved; for the subdivided ecosystems, an overall accuracy of over 88% was achieved, which shows that the present invention takes into account both high overall accuracy and high subject granularity.

[0020] (4) The present invention has a modular structure. The classification hierarchy, rule modules and model combinations can be flexibly replaced according to different regions and target tasks, which makes it easy to migrate and apply to other landscape fragmentation, rapid urbanization or wetland complex areas. Attached Figure Description

[0021] Figure 1 This is a schematic diagram showing the location of the application research area in this embodiment; Figure 2 This is a schematic diagram of the training sample generation method in this embodiment; Figure 3 This is a schematic diagram of the hierarchical classification structure in this embodiment; Figure 4 This is a schematic diagram of the construction based on the spatiotemporal features of multi-source remote sensing data in this embodiment; Figure 5 This is a schematic diagram of the fusion model method of hierarchical classification strategy and probability ensemble method in this embodiment; Figure 6 This is a schematic diagram illustrating the method for integrating pixel scale features, object scale features, and knowledge rules in this embodiment. Figure 7 This is a comparison chart of the accuracy of different models used in this embodiment; Figure 8 This is a comparative analysis chart of spatial recognition capabilities between this embodiment and existing products; Figure 9 This is a comparative analysis chart of the shrub identification capabilities of this embodiment and existing products; Figure 10 This is a comparative analysis chart of wetland identification capabilities between this embodiment and existing products; Figure 11 This is an ecosystem diagram for areas with high-intensity human activity in this embodiment. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0025] Example like Figure 1 As shown, the study area in this embodiment is the Yangtze River Delta region, located in the delta plain formed by alluvial deposits before the Yangtze River flows into the sea. Geographically, it ranges from 114.56°–124.25°E and 26.57°–35.67°N, covering a total area of ​​approximately 358,000 square kilometers, including Shanghai, Jiangsu Province, Zhejiang Province, and Anhui Province. This region has a subtropical monsoon climate, with an average annual precipitation of approximately 1000–1400 mm and an average annual temperature of approximately 14–18 ℃. It also features high-intensity human activity, significant landscape fragmentation, and complex ecosystem types, making it a typical application area for detailed ecosystem mapping methods.

[0026] This embodiment employs a hierarchical modular ecosystem mapping framework construction method for areas with high-intensity human activity, including the following steps: Step S1: Construct a training sample library suitable for large-scale ecosystem mapping by combining automatically generated training samples with visually interpreted samples. Step S1.1: Input multi-source reference data (multi-source land cover or land use product data) for mapping ecosystems in areas of high-intensity human activity. Unify the spatial resolution and coordinate system of the multi-source reference data. Then, based on stable pixels with consistent classification across reference years, multi-product cross-consistency, and neighborhood filtering results, identify candidate sample areas. For example, for main types such as forests, paddy fields, dry land, and urban land, multi-product cross-consistency and neighborhood filtering are used to determine potential areas. Training samples are automatically generated based on the candidate sample areas. Figure 2 As shown in this embodiment, multiple land cover products and thematic data are unified to a spatial resolution of 30 m and a unified coordinate system, and stable pixels with consistent classification in each reference year are selected as potential sample sources.

[0027] Step S1.2: For areas in the candidate sample region that do not meet the conditions for automatic sample generation, visual interpretation is performed based on remote sensing images and their temporal index features. For rare types such as shrubs, grasslands, swamps, flooded land, and bare land, manual interpretation and supplementation are performed by combining high-resolution Google Earth images and the 2019-2021 Landsat NDVI / NDWI time series smoothed by HANTS, and then the target interpretation sample is formed.

[0028] It should be noted that the remote sensing images used for visual interpretation can be Landsat images, Sentinel-2 images, high-resolution remote sensing images, and Gaofen series satellite images. The temporal index features are the NDVI temporal features and NDWI temporal features obtained after temporal smoothing.

[0029] Step S1.3: Based on automatically generated samples and target interpretation samples, construct a training sample library suitable for large-scale ecosystem mapping.

[0030] Step S2: Construct a hierarchical classification structure that conforms to the cognitive rules of ecosystem types, and decompose the current ecosystem mapping task into multiple interconnected sub-tasks. It includes multiple levels L1 to Li, i=2,3,4,... Starting from level L1, all ecosystem types in the target area are initially classified, and then level L2 classification is carried out based on the classification results of level L1. This process is repeated for each level, so as to finally generate high-precision initial ecosystem mapping results.

[0031] like Figure 3As shown, to support ecosystem mapping with a hierarchical structure and strong interpretability, based on basic ecological principles and key remote sensing features, this embodiment constructs a hierarchical classification structure that conforms to the cognitive rules of ecosystem types. This structure includes five levels, L1, L2, L3, L4, and L5, in that order. Level L1 includes permanent water, permanent dry water, and periodic water categories; Level L2 includes vegetation cover and no vegetation cover categories refined from the permanent dry water and periodic water categories in Level L1; Level L3 includes woody vegetation and herbaceous vegetation categories refined from the vegetation cover category in Level L2 corresponding to the permanent dry water category in Level L1; and Level L4 includes water body categories corresponding to the permanent water category in Level L1. Level L1 includes woody vegetation categories (woodland and shrubland), herbaceous vegetation categories (grassland and dryland), and the category of no vegetation cover (built-up land and bare land) derived from the category of permanent waterlessness in Level L2. It also includes paddy fields and marshes derived from the vegetation cover category of periodic water in Level L2, and flooded ponds resulting from the category of no vegetation cover corresponding to the periodic water category in Level L2. Level L5 includes rivers, lakes, and oceans derived from the water body category of Level L4; evergreen forests and deciduous forests derived from the woodland category; inland marshes and coastal marshes derived from the marshland category; and tidal flats and mudflats derived from the flooded pond category. The construction logic is as follows: In Level L1, based on the temporal variation characteristics of surface water, i.e., the duration and frequency of surface water cover, ecosystem types are first divided into three categories: permanent water bodies, non-water bodies, and periodic water bodies. At level L2, based on quantitative vegetation cover thresholds, the major categories are further subdivided into vegetation and non-vegetation subcategories. At levels L3 and L4, these subcategories are further refined based on two key dimensions: dominant vegetation type (e.g., woody vegetation vs. herbaceous vegetation) and intensity of human disturbance (e.g., paddy fields vs. dry land in farmland-related subcategories). Through this refinement process, 10 major ecosystem types are ultimately formed. Finally, at level L5, the four ecologically significant categories—water bodies, forest land, marshland, and flooded land—are further subdivided into more refined types. This refinement is based on a comprehensive consideration of phenological characteristics (e.g., seasonal phenological characteristics of vegetation), morphological attributes (e.g., linear vs. areal characteristics of water bodies), and spatial pattern characteristics (e.g., coastal vs. inland types of marshes).

[0032] It should also be noted that, as Figure 3 The hierarchical modular ecosystem classification system shown is the general organizational framework proposed in this invention. Its hierarchical division approach and modular design principles have good scalability. When applied to other regions, the classification hierarchy, the degree of category subdivision, and the configuration of functional modules can be supplemented, deleted, or reorganized according to the ecosystem composition characteristics, landscape pattern differences, and actual mapping needs of the target region, thereby improving the regional applicability of the method.

[0033] Step S3: Acquire multi-source remote sensing data and construct spatiotemporal features: like Figure 4 As shown, HLSL30 and HLSS30 data, Sentinel-1 dual-polarization data (VV, VH), and Copernicus DEM data for 2020 were acquired. Invalid observations of clouds, cloud shadows, and snow and ice were removed from the HLS data, and a monthly median composite was used to generate a time series. For missing months, linear interpolation was used to reconstruct continuous phenological characteristics. Let a certain band i be in adjacent months... and The effective observations at are respectively and Then the interpolation result for month t can be expressed as:

[0034] in, This represents the interpolation result for band i at month t; and They represent the months. and The valid observations obtained at the location; t represents the month to be estimated; , These are the adjacent valid observation months before and after the month to be estimated. If only one valid observation exists, that valid observation value is used directly as the substitute.

[0035] Based on this, the following features are constructed: 1) Phenological features of the five optical bands of HLS and the indices NDVI, EVI, SAVI, LSWI, NBR, NBR2, NDWI, and MNDWI; 2) Polarization features of SAR polarization quantiles and their interval mean; 3) Maximum value features extracted based on NDVImax and MNDWImax; 4) Temporal features of the annual standard deviation of each optical band, spectral index, and SAR polarization feature; 5) Texture features extracted from near-infrared band images after processing by quantile statistics; 6) Topographic features of elevation, slope, and aspect.

[0036] The formulas for calculating the above indices are as follows: ;

[0037] In the formula, N represents the surface reflectance in the near-infrared band, R represents the surface reflectance in the red band, B represents the surface reflectance in the blue band, G represents the surface reflectance in the green band, S1 represents the surface reflectance in the shortwave infrared 1 band, and S2 represents the surface reflectance in the shortwave infrared 2 band; NDVI represents the normalized vegetation index, EVI represents the enhanced vegetation index, SAVI represents the soil-modified vegetation index, NDWI represents the normalized water index, MNDWI represents the improved normalized water index, LSWI represents the land surface water index, NBR represents the normalized flammability ratio, NBR2 represents the second normalized flammability ratio, 2.5, 6, 7.5 and 1 in the EVI formula are empirical coefficients and constants, respectively, and 0.5 in the SAVI formula is the soil brightness correction factor.

[0038] It should also be noted that, as Figure 4 The spatiotemporal feature construction method based on multi-source remote sensing data shown is a preferred embodiment of this invention. The data source type, time-series synthesis method, feature extraction window, and combination of feature variables used are not fixed. When this method is applied to other regions, the input data source, time-series features, spectral indices, texture features, topographic features, and other auxiliary variables can be added, removed, replaced, or optimized according to the target region's ecosystem composition, phenological rhythm differences, remote sensing data availability, cloud cover conditions, and mapping task requirements, in order to enhance feature representation capabilities and method applicability.

[0039] Step S4: By integrating hierarchical classification strategies and probabilistic ensemble methods, an initial ecosystem spatial distribution map is generated at each level node by combining the outputs of multiple classification models. Step S4.1, as follows Figure 5 As shown, this step can be implemented using a cloud computing platform (such as GEE). Before classification, all input feature data are first normalized to the range [0,1] to ensure the consistency of multi-source feature scales and eliminate model bias introduced by high-amplitude variables. Subsequently, three complementary machine learning base models for feature discrimination are constructed and configured: a Support Vector Machine (SVM) model with a Radial Basis Function (RBF) kernel and parameters set to gamma=0.25 and cost=128; a Random Forest (RF) model with 200 decision trees to balance model stability and generalization performance; and a Gradient Boosting Tree (GTB) model with parameters set to 160 trees, a shrinkage rate of 0.1, a sampling rate of 0.75, and a maximum tree depth of 10, thereby achieving a better balance between classification performance and computational efficiency.

[0040] It should be noted that the aforementioned machine learning base models include the Support Vector Machine (SVM) model, the Random Forest (RF) model which balances model stability and generalization performance, and the Gradient Boosting Tree (GTB), XGBoost, or LightGBM model which achieves a good balance between classification performance and computational efficiency. This embodiment uses the optimal parameter combination for the Yangtze River Delta region's ecosystem mapping task to achieve better classification accuracy and stability. When this method is applied to other regions, the relevant hyperparameters are not fixed but can be adaptively adjusted and optimized based on the target region's ecosystem type composition, landscape fragmentation level, sample quantity and quality, and differences in input feature distribution, through methods such as cross-validation, grid search, or empirical iteration.

[0041] Step S4.2: Given the varying sensitivities of different base classifiers to specific categories, this embodiment introduces a probabilistic ensemble method (PEM) to reduce the uncertainty of single-model predictions and improve the reliability of overall classification. Specifically, based on the trained SVM, RF, and GTB models, a category probability distribution map for each pixel in the image to be classified is generated. Then, the average probability value of the same pixel belonging to each category under each model is calculated. At any classification node at any level, the category label with the highest average probability value is assigned to the corresponding pixel as the final prediction result for that node. The entire classification process strictly follows a hierarchical classification structure, generating an initial ecosystem spatial distribution map through progressive subdivision.

[0042] It should be noted that in the process of calculating the probability value of the same pixel belonging to each category under each model, this embodiment uses the equally weighted average probability value. However, in other application scenarios, a weighted fusion probability value based on the cross-validation accuracy can also be set for different machine learning base models.

[0043] Step S5, as follows Figure 6 As shown, by fusing pixel-scale features, object-scale features, and knowledge rule optimization methods, the initial ecosystem spatial distribution map is refined and corrected, resulting in the following: Figure 11 The ecosystem map shown (uniformly named YRD_EM in this embodiment, hereinafter the same) for areas with high-intensity human activity is as follows: (1) Forest subclass classification: Evergreen forests and deciduous forests are distinguished based on the seasonal NDVI difference (NDVI_max - NDVI_wintermax). In the example, 0.25 is used as the threshold.

[0044] (2) Delineation of wetlands and coastal waters: Coastline data was introduced and a 10 km inland buffer zone was constructed to distinguish between coastal wetlands / coastal waters and inland wetlands / inland waters.

[0045] (3) River and lake separation: Inland water bodies are converted into vector objects, and geometric features such as compactness, density and rectangle fit are extracted at the object level. River and lake separation is achieved based on combination rules. In the embodiment, a combination threshold of "compactness ≤ 16.56, density ≥ 0.65, rectangle fit ≥ 0.21" is used.

[0046] It should also be noted that the threshold parameters in the aforementioned knowledge rules, such as the evergreen forest / deciduous forest classification threshold and the river / lake object discrimination threshold, are all empirical thresholds determined in this embodiment based on regional ecological characteristics and validation samples. When migrating to other regions, the relevant thresholds can be recalibrated according to local vegetation phenological differences, water body morphology characteristics, and geographical background to ensure the applicability of the rule constraints and the reliability of the classification results.

[0047] Example of effect The above embodiments are used to obtain an ecosystem map for areas with high-intensity human activity.

[0048] like Figure 7 The figure shows the accuracy comparison results obtained by using different models in this embodiment: by using the "hierarchical classification strategy + probability ensemble method" for joint modeling, the overall accuracy and stability of the classification of ecosystems in complex areas are significantly improved. The figure clearly shows that compared with the single model method, the overall accuracy is improved by 2.50-5.41 percentage points and the Kappa coefficient is improved by 0.0301-0.0654.

[0049] like Figure 8 The image shown is a comparative analysis of the spatial recognition capabilities of YRD_EM and the existing product (GLC_FCS30D) in this embodiment; as shown... Figure 9 The image shown is a comparative analysis of the shrub recognition capabilities of YRD_EM and the existing product (GLC_FCS30D) in this embodiment; as shown... Figure 10 The figure shown is a comparison analysis of the wetland identification capabilities of YRD_EM and the existing product (GLC_FCS30D) in this embodiment.

[0050] Depend on Figure 8-10 It can be seen that by introducing refined rules based on phenological, morphological, and locational knowledge, the ability to distinguish between subcategories such as rivers / lakes, evergreen forests / deciduous forests, and inland / coastal wetlands has been enhanced, thus making up for the problem that pure spectral classification is insufficient in expressing differences in ecological functions.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for constructing a hierarchical modular ecosystem mapping framework for areas with high-intensity human activity, characterized in that, Includes the following steps: Step S1: Construct a training sample library suitable for large-scale ecosystem mapping by combining automatically generated training samples with visual interpretation samples; Step S2: Construct a hierarchical classification structure that conforms to the cognitive rules of ecosystem types, and decompose the current ecosystem mapping task into multiple interconnected sub-tasks. Step S3: Acquire multi-source remote sensing data and construct spatiotemporal features; Step S4: By integrating hierarchical classification strategies and probabilistic ensemble methods, the initial ecosystem spatial distribution map is generated by combining the outputs of multiple classification models at each level node; Step S5: By fusing pixel-scale features, object-scale features, and knowledge rule optimization methods, the initial ecosystem spatial distribution map is refined and corrected to obtain the final ecosystem map oriented towards areas with high-intensity human activity.

2. The method for constructing a hierarchical modular ecosystem mapping framework for high-intensity human activity areas according to claim 1, characterized in that, Step S1 includes the following steps: Step S1.1: Input multi-source reference data for mapping ecosystems in areas with high-intensity human activity, unify the spatial resolution and coordinate system of the multi-source reference data, and then identify candidate sample regions based on stable pixels with consistent classification in each reference year, cross-consistency of multiple products, and neighborhood filtering results, and automatically generate training samples based on the candidate sample regions. Step S1.2: For regions in the candidate sample region that do not meet the automatic sample generation conditions, perform visual interpretation based on remote sensing images and their temporal index features to form target interpretation samples; Step S1.3: Based on automatically generated samples and target interpretation samples, construct a training sample library suitable for large-scale ecosystem mapping.

3. The method for constructing a hierarchical modular ecosystem mapping framework for high-intensity human activity areas according to claim 1, characterized in that, In step S1.1, the multi-source reference data for mapping ecosystems in areas with high human activity is multi-source land cover or land use product data; in step S1.2, the remote sensing images used for visual interpretation can be Landsat images, Sentinel-2 images, high-resolution remote sensing images, or Gaofen series satellite images, and the temporal index features are NDVI temporal features and NDWI temporal features obtained after temporal smoothing processing.

4. The method for constructing a hierarchical modular ecosystem mapping framework for areas with high-intensity human activity, as described in claim 1, is characterized in that... In step S2, the hierarchical classification structure that conforms to the cognitive rules of ecosystem types includes multiple levels L1~Li, i=2,3,4,... Starting from level L1, all ecosystem types in the target area are initially classified, and then level L2 classification is carried out based on the classification results of level L1. Then, each level is like this, so as to finally generate a high-precision initial ecosystem mapping result.

5. The method for constructing a hierarchical modular ecosystem mapping framework for areas with high-intensity human activity, as described in claim 4, is characterized in that... In step S2, the hierarchical classification structure conforming to the cognitive rules of ecosystem types includes five levels, namely L1, L2, L3, L4, and L5. Level L1 includes permanent water, permanent dry water, and periodic water categories. Level L2 includes vegetation cover and no vegetation cover categories refined from the permanent dry water and periodic water categories in Level L1. Level L3 includes woody vegetation and herbaceous vegetation categories refined from the vegetation cover category in Level L2 corresponding to the permanent dry water category in Level L1. Level L4 includes the water body category corresponding to the permanent water category in Level L1, and the woody vegetation category refined from the woody vegetation category in Level L3, including woodland and shrubland. The categories include grassland and dryland categories refined from the herbaceous vegetation category in level L3; built-up land and bare land categories refined from the no-vegetation-cover category corresponding to the permanent waterless category in level L2; paddy fields and marshes categories refined from the vegetation-cover category corresponding to the periodic water category in level L2; and flooded ponds categories generated by the no-vegetation-cover category corresponding to the periodic water category in level L2. Level L5 includes rivers, lakes, and ocean categories refined from the water body category in level L4; evergreen forest and deciduous forest categories refined from the forest category; inland marshes and coastal marshes categories refined from the marshland category; and tidal flats and mudflats categories refined from the flooded pond category.

6. The method for constructing a hierarchical modular ecosystem mapping framework for high-intensity human activity areas according to claim 1, characterized in that, Step S3 specifically involves acquiring multi-source optical remote sensing data, Sentinel-1 dual-polarization data, and digital elevation model data for the required year; preprocessing the multi-source optical remote sensing data to remove invalid observations affected by clouds, cloud shadows, and snow and ice; and constructing a time series using monthly median synthesis. Missing months are reconstructed using linear interpolation; spectral index features, maximum value features, temporal features, and texture features are extracted based on the reconstructed optical time series; polarization features are extracted based on the Sentinel-1 dual-polarization data; and terrain features are extracted based on the digital elevation model data. The multi-source optical remote sensing data includes HLSL30 data and HLSS30 data. The spectral index features include one or more of NDVI, EVI, SAVI, NDWI, MNDWI, LSWI, NBR, and NBR2 calculated based on the reconstructed optical time series. Missing months are reconstructed using linear interpolation.

7. The method for constructing a hierarchical modular ecosystem mapping framework for areas with high-intensity human activity, as described in claim 4, is characterized in that... Step S4 includes the following steps: Step S4.1: Before performing classification, the spatiotemporal features constructed in step S3 are uniformly normalized to the range of [0,1]. Then, a machine learning base model with complementary features in feature discrimination is constructed and configured. Step S4.2: Based on the trained machine learning base model, generate the category probability distribution map of each pixel in the image to be classified, and then calculate the probability value of the same pixel belonging to each category under each model. At any level of classification node, assign the category with the highest probability value to the corresponding pixel as the final prediction result of that node. The entire classification process strictly follows the hierarchical structure classification. Through step-by-step subdivision, the initial ecosystem spatial distribution map is finally generated.

8. The method for constructing a hierarchical modular ecosystem mapping framework for high-intensity human activity areas according to claim 5, characterized in that, Step S5 includes the following steps: Step S5.1: Based on the seasonal NDVI difference, a preset threshold is used to distinguish between evergreen forests and deciduous forests; Step S5.2: Based on the coastline data, construct an inland buffer zone and distinguish between coastal wetlands / coastal water bodies and inland wetlands / inland water bodies; Step S5.3: Convert inland water bodies into vector objects, extract geometric features such as compactness, density, and rectangle fit at the object level, and achieve river-lake separation based on combination rules.

9. The method for constructing a hierarchical modular ecosystem mapping framework for areas with high-intensity human activity, as described in claim 7, is characterized in that... In step S4.1, the machine learning base models are the Support Vector Machine (SVM) model, the Random Forest (RF) model which balances model stability and generalization performance, and the Gradient Boosting Tree (GPB) model which achieves a better balance between classification performance and computational efficiency. The GPB model can be GTB, XGBoost, or LightGBM.

10. The method for constructing a hierarchical modular ecosystem mapping framework for high-intensity human activity areas according to claim 7, characterized in that, In step S4.2, when calculating the probability value of the same pixel belonging to each category under each model, the probability value can be the equally weighted average probability value, or a weighted fusion probability value based on the cross-validation accuracy can be set for different machine learning base models.

Images