A natural ecological element classification method and system based on an adaptive deep forest model

Abstract

This invention discloses a natural ecological element classification method and system based on an adaptive deep forest model, relating to the field of remote sensing monitoring technology. The method includes acquiring classification element data and multi-source remote sensing data of the area to be interpreted; preprocessing the multi-source remote sensing data to obtain multi-source standard data; constructing a natural element classification system and using the classification element data to construct a classification feature dataset; constructing a local reference surface for mountain discrimination and minimum patch cleaning to obtain mountain extraction results; inputting the classification feature dataset into a trained adaptive deep forest model for classification interpretation to obtain a natural element classification result layer; and overlaying these layers to obtain the final natural element classification result. This invention, by constructing a highly adaptable classification framework, systematically divides and integrates eight categories of ecological elements—mountains, water, forests, fields, lakes, grasslands, sand, and ice—to more accurately reflect the overall pattern of the ecosystem.

Description

Technical Field

[0001] This invention relates to the field of remote sensing monitoring technology, and in particular to a method and system for classifying natural ecological elements based on an adaptive deep forest model. Background Technology

[0002] Mountains, rivers, forests, fields, lakes, grasslands, sand, and ice are the core elements constituting a natural ecosystem. These elements are closely interdependent and interact with each other, forming an inseparable organic community of life. Different ecological elements are spatially heterogeneous and intertwine and merge, collectively creating a complex ecological pattern. To deeply understand the spatial intergrowth relationships among these elements and the distribution characteristics of the ecological pattern, it is necessary to establish a direct and highly adaptable classification system framework. Traditional land use classification systems often suffer from strong fragmentation and fail to effectively reflect the connections between ecological elements, while the fragmentation of each element also limits the holistic representation of the ecosystem.

[0003] Remote sensing technology, as a crucial means of acquiring land cover information, has been widely applied in natural resource surveys, land use classification, and ecological environment monitoring. Traditional remote sensing classification methods mainly include pixel-based supervised classification, unsupervised classification, and object-based image analysis methods, which can effectively identify typical land cover types. With the development of machine learning and deep learning technologies, methods such as support vector machines, random forests, and convolutional neural networks have been gradually introduced into the field of remote sensing classification, significantly improving classification accuracy and applicability. In recent years, the fusion of multi-source remote sensing data (such as optical imagery, radar data, and high-resolution satellite imagery) has further promoted the refined interpretation of land cover information, resulting in significant improvements in the accuracy and applicability of remote sensing classification methods.

[0004] While existing remote sensing classification methods have made significant progress in terms of technical accuracy and applicability, some limitations remain. Traditional remote sensing monitoring methods typically rely on single features such as spectral or textural characteristics, making it difficult to comprehensively reflect the complex structure and interrelationships of multiple elements in an ecosystem. This results in certain deficiencies in the classification results when expressing ecological patterns. Even with the adoption of machine learning and deep learning methods, which have further improved classification accuracy and robustness, most studies still rely on traditional land use / cover classification frameworks. These frameworks tend to favor single land types such as cultivated land, forest land, water bodies, and construction land. They not only fail to comprehensively characterize the integrity and interactions of ecological community elements such as mountains, rivers, forests, fields, lakes, grasslands, sand, and ice, but also struggle to comprehensively and systematically express the multidimensional characteristics of spatial heterogeneity and interactions among these complex ecological elements. Particularly in arid and desert regions, existing methods are even less able to fully reflect the spatial heterogeneity of ecological patterns and the interactions between multiple elements, failing to meet the needs of ecological civilization construction and sustainable development strategies for comprehensive ecological monitoring.

[0005] Therefore, how to construct a classification system that can meet the needs of the ecological community, effectively represent complex ecological elements and their interrelationships, and be compatible with and accurately reflect the classification framework of the eight major elements of mountains, rivers, forests, fields, lakes, grasslands, sand, and ice, so as to provide more reliable and effective support for the division of ecological units, the evaluation of patterns, and the comprehensive management of ecosystems, has become a key technical issue in promoting the division of ecological units and the analysis of patterns. Summary of the Invention

[0006] In view of this, the present invention provides a natural ecological element classification method and system based on an adaptive deep forest model. By constructing a highly adaptable classification framework and combining multi-source remote sensing data and machine learning models, it systematically classifies and integrates eight ecological elements—mountains, water, forests, fields, lakes, grasslands, sand, and ice—to more accurately reflect the overall pattern of the ecosystem. At the same time, the present invention also significantly improves the applicability of the ecological element classification results in ecological unit division, pattern evaluation, and system governance, providing strong technical support for multi-regional and cross-scale ecological monitoring.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for classifying natural ecological elements based on an adaptive deep forest model includes the following steps: S1 Acquisition Steps: Acquire classification feature data and multi-source remote sensing data of the area to be interpreted; S2 processing steps: Use affine transformation and resampling association techniques to preprocess multi-source remote sensing data to obtain multi-source standard data; S3 system setup steps: Construct a natural element classification system based on traditional land use classification; S4 Feature Construction Steps: Based on the natural element classification system, construct a classification feature dataset using classification element data; the classification feature dataset does not include mountain element features; S5 Mountain Extraction Steps: Construct a local reference surface, use slope constraints and morphology to perform mountain discrimination and minimum patch cleaning on multi-source standard data, and obtain the mountain extraction results; S6 model classification steps: Input the classification feature dataset into the trained adaptive deep forest model to classify and interpret the natural elements, and obtain the natural element classification result layer; S7 Overlay Output Steps: The natural element classification result layer and the mountain extraction result are overlaid through spatial analysis to obtain the final natural element classification result of the area to be interpreted.

[0008] Optionally, the multi-source remote sensing data in S1 includes: 30m land cover data of the region to be interpreted obtained through GLC_FCS30, 30m land cover data obtained through CLCD, 10m land cover data obtained through ESA, 500m land cover data obtained through MODIS, and 10m land cover data obtained through Xinjiang regional land cover classification data, where m represents resolution.

[0009] Optionally, the preprocessing in S2 of the above method includes: performing radiometric correction, atmospheric correction, and geometric correction on the multi-source remote sensing data, as well as spatial registration and resolution normalization processing based on affine transformation and resampling association techniques to obtain multi-source standard data of the region to be interpreted.

[0010] The above method can be optionally used. The natural element classification system in S3 is an eight-element classification system of "mountains, water, forests, fields, lakes, grasslands, sand, and ice". Specifically, it includes: mountain element classification based on topographic relief areas; water element classification including linear and planar water bodies; forest element classification including natural and artificial forests; field element classification including agricultural cultivated land and multiple cropping plots; lake element classification for independent water bodies; grassland element classification including natural grasslands, pastures, and low grass and shrub vegetation; sand element classification; and ice element classification.

[0011] The above method, optionally, includes S5 specifically: S51 Measurement Steps: Construct a local reference surface, determine the regional relative undulation of multi-source standard data and measure the uplift, and determine the relative undulation value; S52 discrimination steps: By using slope constraints and relative undulation thresholds, mountain discrimination is performed on multi-source standard data to determine the threshold discrimination rules; S53 Cleaning Steps: Morphology is introduced, and threshold discrimination rules are used to perform minimum spot cleaning on multi-source standard data to obtain the mountain extraction results.

[0012] Optionally, the formula for calculating the datum plane in S51 is as follows: ； The formula for calculating relative fluctuations is: ; In the formula, Indicates the original DEM elevation. y Represents a pixel. Represented by pixels x The central area p Indicates the lower quantile. Represents a cell x The original DEM elevation, Represents a pixel xThe elevation of the reference surface.

[0013] Alternatively, the formula for calculating the slope constraint in S52, as described above, is: ; The formula for calculating the threshold discrimination rule is: ; In the formula, Indicates the original DEM elevation. x , y These represent the current cell and the cells in its neighborhood, respectively. express Preliminary judgment results R(x) Indicates relative fluctuation value, Indicates the relative fluctuation threshold. This is the slope threshold.

[0014] The above method, optionally, includes the following specific steps in the training process of the adaptive deep forest model in S6: Acquisition steps: Obtain the classification feature data for training and divide it into training set and validation set; Scanning steps: Perform multi-granularity feature scanning on the training set to obtain multi-dimensional feature vectors; Input steps: Input the multi-dimensional feature vector into the first layer of the cascaded forest structure random forest model for feature concatenation to obtain the first prediction vector; Training steps: For any layer after the first layer of the random forest model, the prediction vector of the previous layer is used as input to perform feature concatenation to obtain the prediction vector of the current layer, and the training result of the current layer is determined based on the validation set. Verification steps: Compare the training results of the current layer with the training results of the previous layer to obtain the performance improvement results; if the performance improvement results meet the preset threshold, then determine the current layer random forest model as the last layer random forest model and execute the next step; if the performance improvement results do not meet the preset threshold, then repeat the training steps. The steps are as follows: Based on the sparsity screening of logistic regression, invalid models from the first layer to the last layer of the random forest model are removed by Elastic Net regularization to obtain a trained adaptive deep forest model.

[0015] The above method, optionally, uses the following formula to calculate the objective function of Elastic Net regularization: ; in, This represents the L1 norm, used to generate sparse solutions; λ represents the L2 norm, used to generate stable solutions; λ is the regularization strength, and λ≥0; ∈[0,1], used to control the ratio of L1 and L2.

[0016] A natural ecological element classification system based on an adaptive deep forest model, used to implement the natural ecological element classification method based on an adaptive deep forest model as described above, includes a data acquisition module, a processing module, a system setting module, a feature construction module, a mountain extraction module, a model classification module, and an overlay output module connected in sequence. The acquisition module is used to acquire classification feature data and multi-source remote sensing data of the area to be interpreted; The processing module is used to preprocess multi-source remote sensing data using affine transformation and resampling correlation techniques to obtain multi-source standard data; The system setting module is used to construct a classification system for natural elements based on traditional land use classification. The feature construction module is used to construct a classification feature dataset based on the natural element classification system and using classification element data; the classification feature dataset does not include mountain element features; The mountain extraction module is used to construct a local reference surface, and to perform mountain discrimination and minimum patch cleaning on multi-source standard data using slope constraints and morphology to obtain mountain extraction results. The model classification module is used to input the classification feature dataset into the trained adaptive deep forest model to classify and interpret natural elements, and obtain a natural element classification result layer. The overlay output module is used to overlay the natural element classification result layer and the mountain extraction result through spatial analysis to obtain the final natural element classification result of the area to be interpreted.

[0017] As can be seen from the above technical solution, compared with the prior art, the present invention provides a method and system for classifying natural ecological elements based on an adaptive deep forest model, which has the following beneficial effects: (1) Improve the accuracy of ecological element identification and classification: This invention innovatively proposes an eight-element classification system of "mountains, rivers, forests, fields, lakes, grasslands, sand and ice". Combined with multi-source remote sensing data, it can comprehensively and systematically reflect the spatial relationship and interaction between different ecological elements. Through multi-dimensional information fusion, it overcomes the problem of insufficient classification accuracy of traditional methods under complex ecological patterns and significantly improves the accuracy and adaptability of surface ecosystem classification. (2) Improved accuracy and stability of mountain extraction: The mountain extraction technology proposed in this application achieves high-precision mountain area identification by combining local datum surface calculation, slope constraint and morphological processing; and can effectively identify complex and significantly changing terrain, especially suitable for mountainous and marginal areas, thus improving the accuracy and stability of mountain extraction and providing reliable basic data for subsequent ecological analysis and resource management. (3) Improving the classification accuracy and robustness of remote sensing data: This application introduces an adaptive deep forest model (ADeFS) and combines it with multi-source remote sensing data (such as optical images, radar data, topographic data, etc.) to effectively optimize feature selection and shrinkage techniques, thereby enhancing the model's ability to identify complex ecological elements. This model not only improves classification accuracy but also enhances adaptability to complex regions, especially showing stronger robustness and generalization ability in arid and desert regions; (4) The method and system disclosed in this invention can be compatible with and accurately reflect the classification framework of the eight elements of mountains, rivers, forests, fields, lakes, grasslands, sand and ice, providing more reliable and effective support for the division of ecological units, the evaluation of patterns and the comprehensive management of ecosystems. It also significantly improves the applicability of the ecological element classification results in the division of ecological units, the evaluation of patterns and the management of systems, and provides strong technical support for multi-regional and cross-scale ecological monitoring. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This invention discloses a flowchart of a natural ecological element classification method based on an adaptive deep forest model; Figure 2 This is a schematic diagram of multi-source remote sensing data disclosed in an embodiment of the present invention; Figure 3 This is the mountain extraction result of the region to be interpreted obtained from multi-source standard data, as disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of the module structure of the Adaptive Deep Forest Model (ADeFS) disclosed in an embodiment of the present invention; Figure 5 This is a schematic diagram of the final natural element classification result obtained by spatial analysis and overlay as disclosed in an embodiment of the present invention. Figure 6 This is a schematic diagram showing the coefficient evaluation results and accuracy of the trained adaptive deep forest model disclosed in an embodiment of the present invention. Detailed Implementation

[0020] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0022] See Figure 1 As shown, this invention discloses a method for classifying natural ecological elements based on an adaptive deep forest model, comprising the following steps: S1 Acquisition Steps: Acquire classification feature data and multi-source remote sensing data of the area to be interpreted; S2 processing steps: Use affine transformation and resampling association techniques to preprocess multi-source remote sensing data to obtain multi-source standard data; S3 system setup steps: Construct a natural element classification system based on traditional land use classification; S4 Feature Construction Steps: Based on the natural element classification system, construct a classification feature dataset using classification element data; the classification feature dataset does not include mountain element features; S5 Mountain Extraction Steps: Construct a local reference surface, use slope constraints and morphology to perform mountain discrimination and minimum patch cleaning on multi-source standard data, and obtain the mountain extraction results; S6 model classification steps: Input the classification feature dataset into the trained adaptive deep forest model to classify and interpret the natural elements, and obtain the natural element classification result layer; S7 Overlay Output Steps: The natural element classification result layer and the mountain extraction result are overlaid through spatial analysis to obtain the final natural element classification result of the area to be interpreted.

[0023] To enable the present invention Figure 1The disclosed technical solution is more specific. The following is a detailed explanation of the steps of a natural ecological element classification method based on an adaptive deep forest model.

[0024] Optionally, the classification element data in S1 includes optical imagery data, vegetation and water indices, radar information data, and topographic factors.

[0025] Referring to Table 1, the optical image data are optical images acquired by Landsat 8 (formerly known as LDCM, Landsat Data Continuity Mission, the eighth land observation satellite). Landsat 8 extracts surface reflectance data of bands 1–10 of the area to be interpreted to reveal the spectral differences and energy characteristics of different ecological elements.

[0026] The vegetation and water indices are calculated based on Landsat 8 optical data. The vegetation and water indices include NDVI (Normalized Difference Vegetation Index), NDWI (Normalized Difference Water Index), and SAVI (Soil Adjusted Vegetation Index). NDVI is used to characterize vegetation cover and vitality; NDWI is used for water body identification and wetland monitoring; and SAVI is used to enhance the accuracy of vegetation identification in arid and semi-arid areas.

[0027] The radar information data specifically refers to Sentinel-1 dual-polarization radar data. The backscattering coefficients after VV and VH polarization are extracted, and their ratios are calculated. This radar information data is primarily used to characterize surface morphology, moisture content, and structural features. This type of data effectively compensates for the limitations of optical imagery under cloudy or snow-covered conditions, enhancing the comprehensiveness of the data.

[0028] Topographic factors include absolute elevation obtained from SRTM data, and further calculated slope, aspect, and roughness indices. Topographic factors reflect the constraints of topographic relief on the distribution of ecological elements in the area to be interpreted, and help extract mountain features in the area to be interpreted. SRTM (Shuttle Radar Topography Mission) is a 30m / 90m resolution DEM of the global near-land area (approximately 80%) acquired by NASA and NGA using the Endeavour spacecraft. It is a core foundational data for global topographic / hydrological / ecological research.

[0029] Furthermore, the multi-source remote sensing data in S1 specifically includes: 30m land cover data of the region to be interpreted obtained through GLC_FCS30, 30m land cover data obtained through CLCD, 10m land cover data obtained through ESA, 500m land cover data obtained through MODIS, and 10m land cover data obtained through Xinjiang regional land cover classification data, where m represents resolution.

[0030] This invention obtains land cover data for the area to be interpreted by collecting and organizing five types of data products from the 2020 research demonstration area: GLC_FCS30 (global fine land cover data, 30m), CLCD (China land cover dataset, 30m), ESA (global land cover dataset, 10m), MODIS (MCD12Q1 land cover data, 500m), and Xinjiang regional land cover classification data (10m). Referring to Table 1, the collected and organized land cover data is mapped to the ecological element categories of "water, forest, farmland, lake, grassland, sand, and ice." A weighted voting strategy is used, based on decision-level fusion and weighted voting, to obtain the final basic product for land classification of the research area, namely, multi-source remote sensing data.

[0031] Table 1. Detailed table of weight allocation and rationale for each data product.

[0032] Optionally, the preprocessing in S2 specifically includes: performing radiometric correction, atmospheric correction, and geometric correction on multi-source remote sensing data, as well as spatial registration and resolution normalization processing based on affine transformation and resampling association techniques to obtain multi-source standard data of the region to be interpreted.

[0033] Reference Figure 2 The aforementioned multi-source remote sensing data, after undergoing unified radiometric, atmospheric, and geometric corrections, combined with spatial registration and resolution standardization, forms a multi-dimensional, highly complementary fusion feature dataset, providing reliable input for subsequent classification system construction and model training.

[0034] Optionally, the natural element classification system in S3 is an eight-element classification system of "mountains, waters, forests, fields, lakes, grasslands, sand, and ice". Specifically, it includes: mountain element classification based on topographic relief areas; water element classification including linear and areal water bodies; forest element classification including natural and artificial forests; field element classification including agricultural cultivated land and multi-cropping plots; lake element classification for independent water bodies; grassland element classification including natural grasslands, pastures, and low-lying grass and shrub vegetation; sand element classification; and ice element classification.

[0035] To meet the needs of representing the overall pattern of ecosystems, the embodiments disclosed in this invention, based on traditional land use classification, adopt an eight-element classification system of "mountains, rivers, forests, fields, lakes, grasslands, sand, and ice" oriented towards the ecological community. The specific classification is as follows: Mountains: mainly consisting of undulating terrain areas, including bare land, sandy areas, and mountain vegetation, serving as an embedded layer; Water: refers to linear and planar water bodies such as rivers, reservoirs, and ponds; Forests include both natural forests and plantations; Field: refers to agricultural land and land used for multiple cropping; Lake: A separate body of water, distinct from rivers; Grass: including natural grasslands, pastures, and low-lying grass and shrub vegetation; Sand: Represents deserts, sand dunes, and areas of wind and sand activity; Ice: encompasses glaciers, ice fields, and areas covered by snow all year round.

[0036] The natural element classification system of this invention achieves a systematic division and integrated expression of eight categories of ecological elements: mountains, rivers, forests, fields, lakes, grasslands, sand, and ice. Through this system, this invention can accurately describe the spatial distribution patterns of natural resources and ecological elements within a unified framework, providing standardized support for subsequent feature construction, classification interpretation, and ecological pattern analysis.

[0037] Table 2. Characteristics and Differentiating Functions of Classification Elements

[0038] Referring to Table 2, S4, the embodiments of the present invention are based on the ecological element classification system of "mountains, rivers, forests, fields, lakes, grasslands, sand and ice". Combining multi-dimensional features such as spectrum, index, thermal infrared, radar and topography, a classification feature dataset other than the "mountain" element is constructed. This dataset is not used as feature input, but is only logically superimposed on the final result.

[0039] Optical image data is used to characterize the spectral features of the area to be interpreted, with band reflectance as a specific indicator. Vegetation and non-vegetation are distinguished by red light-near infrared, and snow and desert bare land are identified by SWIR (Short-Wave Infrared). It can also distinguish between forest / grass / field and water / ice / sand.

[0040] Vegetation and water indices are used to characterize the vegetation and water features of the area to be interpreted. NDVI identifies the difference between forest / grassland and sand / bare land, NDWI identifies the difference between water / lake and land, and SAVI identifies the difference between water / lake and land.

[0041] Radar information data is used to describe the radar characteristics of the area to be interpreted. By extracting the backscattering coefficients after VV and VH polarization, the data utilizes the characteristics of low backscattering in water bodies and strong backscattering in forest and grassland areas to characterize the land surface morphology, water content and structural features, and provides a reason for distinguishing between forest / grassland, sand / bare land, water / lake and land.

[0042] Topographic factors are used to describe the topographic features of the area to be interpreted. The specific index is Roughness, which measures the elevation undulation of the surrounding area, thus digitizing the constraints of topographic undulation on the distribution of ecological elements in the area to be interpreted. SRTM (Shuttle Radar Topography Mission) is a 30m / 90m resolution DEM of the global near-land area (approximately 80%) acquired by NASA and NGA using the Endeavour spacecraft. It is a core foundational data for global topographic / hydrological / ecological research.

[0043] Optionally, S5 specifically includes: S51 Measurement Steps: Construct a local reference surface, determine the regional relative undulation of multi-source standard data and measure the uplift, and determine the relative undulation value; S52 discrimination steps: By using slope constraints and relative undulation thresholds, mountain discrimination is performed on multi-source standard data to determine the threshold discrimination rules; S53 Cleaning Steps: Morphology is introduced, and threshold discrimination rules are used to perform minimum spot cleaning on multi-source standard data to obtain the mountain extraction results.

[0044] Reference Figure 3 The invention discloses a method for extracting mountain ranges. In step S5, it extracts the distribution of mountains within a region by establishing regional mountain range extraction principles. The specific extraction rules are as follows: (1) In S51, the embodiment of the present invention determines the relative undulation of the region by constructing a local reference surface in order to measure the amount of uplift of the current pixel relative to the surrounding "low-lying terrain".

[0045] The formula for calculating the datum plane is: ； The formula for calculating relative fluctuations is: ; In the formula, Indicates the original DEM elevation. y Represents a pixel. Represented by pixels x The central area p Indicates the lower quantile. Represents a pixel x The original DEM elevation, Represents a pixel x The elevation of the reference surface.

[0046] (2) Further, in S52, the embodiments of the present invention use slope constraints to identify mountains, satisfy the criteria of "significant elevation difference" and "obvious slope" and include short and steep slopes to improve the recognition accuracy.

[0047] The slope is approximated using the DEM gradient, and the formula for calculating the slope constraint is: ; The formula for calculating the threshold discrimination rule is: ; In the formula, Indicates the original DEM elevation. x , y These represent the current cell and the cells in its neighborhood, respectively. express Preliminary judgment results R(x) Indicates relative fluctuation value, Indicates the relative fluctuation threshold. This is the slope threshold.

[0048] The embodiments of the present invention further incorporate short, steep slopes, calculated using the following formula: .

[0049] (3) Furthermore, in S53 of this embodiment of the invention, morphology is specifically introduced to perform minimum spot cleaning and maintain the control. Figure 1 Desire

[0050] Number of connected cells (8-neighborhood): ; Convert the pixel threshold based on the minimum target area (hectares): ; The extracted results of the cleaned mountain body are represented as follows: .

[0051] Optionally, the training process of the adaptive deep forest model in S6 specifically includes: Acquisition steps: Obtain the classification feature data for training and divide it into training set and validation set; Scanning steps: Perform multi-granularity feature scanning on the training set to obtain multi-dimensional feature vectors; Input steps: Input the multi-dimensional feature vector into the first layer of the cascaded forest structure random forest model for feature concatenation to obtain the first prediction vector; Training steps: For any layer after the first layer of the random forest model, the prediction vector of the previous layer is used as input to perform feature concatenation to obtain the prediction vector of the current layer, and the training result of the current layer is determined based on the validation set. Verification steps: Compare the training results of the current layer with the training results of the previous layer to obtain the performance improvement results; if the performance improvement results meet the preset threshold, then determine the current layer random forest model as the last layer random forest model and execute the next step; if the performance improvement results do not meet the preset threshold, then repeat the training steps. The steps are as follows: Based on the sparsity screening of logistic regression, invalid models from the first layer to the last layer of the random forest model are removed by Elastic Net regularization to obtain a trained adaptive deep forest model.

[0052] Furthermore, the specific formula for calculating the objective function of Elastic Net regularization is as follows: ; in, This represents the L1 norm, used to generate sparse solutions; λ represents the L2 norm, used to generate stable solutions; λ is the regularization strength, and λ≥0; ∈[0,1], used to control the ratio of L1 and L2.

[0053] Specifically, in S6, after completing the classification system setting, feature construction and mountain extraction, the present invention uses the Automatic Deep Forest Shrinkage (ADeFS) model to classify and interpret the seven elements of "water, forest, field, lake, grass, sand and ice".

[0054] ADeFS is a deep ensemble forest model. This invention aims to optimize the structure of the deep forest by introducing a shrinkage technique, reducing redundant trees and inefficient forests, thereby significantly reducing classification complexity while maintaining the accuracy of natural element classification prediction. The training process of this invention, matched with the forest model, consists of four key steps, achieving layer-by-layer reinforcement learning through multi-granularity feature scanning and cascaded structures. Compared to traditional machine learning methods, ADeFS has adaptive feature selection capabilities and strong nonlinear modeling capabilities; compared to deep neural networks, it has a smaller parameter scale, lower requirements for training sample size, and stronger robustness and generalization ability, making it particularly suitable for regional ecological element classification tasks based on multi-source remote sensing data. A schematic diagram of the ADeFS module structure is shown below. Figure 4 As shown, the detailed training steps are mainly divided into: Multi-Grained Scanning (MGS): This technique scans the input classification feature dataset using a seeded random window generation strategy. While maintaining the window length and the number of generated subsamples, it introduces random initial offsets within the feature space to obtain differentiated local perspectives at the same scale. Multiple randomization starting points are used to generate multi-dimensional feature representations, capturing the spatial and spectral diversity of ecological elements. The multi-dimensional feature vectors are then input into Random Forest (RF) and Completely Random Forest (CRF) respectively, and the outputs of all forests are concatenated to form the enhanced feature vector.

[0055] The multi-granularity feature scanning has a window size of 6 and a number of windows of 4. Each window trains 4 forests, and each forest contains 50 trees. The step size is achieved by randomizing the starting point to realize differentiated local perspectives.

[0056] Cascade Forest (CF) architecture utilizes a hierarchical structure composed of multiple random forests and fully random forests to extract and combine information layer by layer, progressively improving the model's classification and discriminative power. CF employs a hierarchical structure. In the nth layer, the model receives the output prediction vector from the n-1th layer and concatenates it with the original input features to form a new input representation. Each layer contains k forests, and each forest contains m decision trees. At the tree level, when a sample falls into a leaf node, the output is the classification task for that leaf node; at the forest level, the outputs of the m trees within that forest are averaged to form the forest-level prediction vector. Then, the outputs of all forests in the same layer are concatenated to obtain the overall representation of that layer, which is used as the input for the next layer. This process iterates, and the model gradually accumulates discriminative information until the performance improvement on the validation set is insufficient to meet a preset threshold, at which point the number of layers is stopped, i.e., adaptive shrinkage (such as an early stopping strategy based on error convergence).

[0057] The preset threshold for improving validation set performance is generally set by setting the performance of four consecutive layers of the model, or stopping when the training score of the current layer does not exceed that of the immediately preceding layer. The specific logic is: the score of the current layer is less than or equal to the best score.

[0058] Shrinkage steps: Refer to Figure 4 After verification, for the large number of redundant forests in the first to last layers of CF, a sparsity screening based on logistic regression is adopted. Elastic Net regularization is used to sparsely model the forest output features, and the L1 and L2 norms of the forest grouping are combined with the objective function for scoring to achieve automatic removal of invalid forests; for example... Figure 4The partial sparsification step in the middle of the process filters out the squares covered by the red horizontal lines, which means removing the low-contribution and redundant parts of the forest model. This achieves a balance between model complexity and classification accuracy, and alleviates the problems of too many trees, overfitting and high computational cost.

[0059] Ensemble: After invalid removal, the selected forests are weighted by normalized weights (weights come from the aforementioned sparse solution / group norm scores), aligning the feature selection weights with the ensemble weights to improve the robustness of the model and obtain the final trained adaptive deep forest model.

[0060] Reference Figure 5 In S7, after inputting the multi-source feature data of the region to be interpreted into the trained adaptive deep forest model, the output layer corresponds to the classification results of 7 ecological elements. Then, through spatial analysis and overlaying of the mountain extraction results, the final result is obtained as shown below. Figure 5 The final natural element classification results for the region to be interpreted are shown.

[0061] This invention further employs three indicators—confusion matrix, overall accuracy, and Kappa coefficient—to evaluate the final natural element classification results, and combines them with manually interpreted data for cross-validation to ensure the reliability and applicability of the classification results. The model results and accuracy are attached. Figure 6 As shown, the overall accuracy is 93.65%, and the Kappa coefficient is 0.9117.

[0062] and Figure 1 Corresponding to the method shown, this invention also discloses a natural ecological element classification system based on an adaptive deep forest model, used to achieve, for example... Figure 1 The natural ecological element classification method shown includes, in sequence, a data acquisition module, a processing module, a system setting module, a feature construction module, a mountain extraction module, a model classification module, and an overlay output module; The acquisition module is used to acquire classification feature data and multi-source remote sensing data of the area to be interpreted; The processing module is used to preprocess multi-source remote sensing data using affine transformation and resampling correlation techniques to obtain multi-source standard data; The system setting module is used to construct a classification system for natural elements based on traditional land use classification; The feature building module is used to construct a classification feature dataset based on the natural element classification system and using classification element data; the classification feature dataset does not include mountain element features; The mountain extraction module is used to construct a local reference surface, and uses slope constraints and morphology to perform mountain discrimination and minimum patch cleaning on multi-source standard data to obtain mountain extraction results. The model classification module is used to input the classification feature dataset into the trained adaptive deep forest model to classify and interpret natural elements, and obtain the natural element classification result layer. The overlay output module is used to overlay the natural element classification result layer and the mountain extraction result through spatial analysis to obtain the final natural element classification result of the area to be interpreted.

[0063] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for classifying natural ecological elements based on an adaptive deep forest model, characterized in that, Includes the following steps: S1 Acquisition Steps: Acquire classification feature data and multi-source remote sensing data of the area to be interpreted; S2 processing steps: Use affine transformation and resampling association techniques to preprocess multi-source remote sensing data to obtain multi-source standard data; S3 system setup steps: Construct a natural element classification system based on traditional land use classification; S4 Feature Construction Steps: Based on the natural element classification system, construct a classification feature dataset using classification element data; The classification feature dataset does not include mountain feature features; S5 Mountain Extraction Steps: Construct a local reference surface, use slope constraints and morphology to perform mountain discrimination and minimum patch cleaning on multi-source standard data, and obtain the mountain extraction results; S6 model classification steps: Input the classification feature dataset into the trained adaptive deep forest model to classify and interpret the natural elements, and obtain the natural element classification result layer; S7 Overlay Output Steps: The natural element classification result layer and the mountain extraction result are overlaid through spatial analysis to obtain the final natural element classification result of the area to be interpreted.

2. The natural ecological element classification method based on an adaptive deep forest model according to claim 1, characterized in that, The multi-source remote sensing data in S1 specifically includes: 30m land cover data of the region to be interpreted obtained through GLC_FCS30, 30m land cover data obtained through CLCD, 10m land cover data obtained through ESA, 500m land cover data obtained through MODIS, and 10m land cover data obtained through Xinjiang regional land cover classification data, where m represents resolution.

3. The natural ecological element classification method based on an adaptive deep forest model according to claim 1, characterized in that, The preprocessing in S2 specifically includes: performing radiometric correction, atmospheric correction, and geometric correction on multi-source remote sensing data, as well as spatial registration and resolution normalization based on affine transformation and resampling correlation techniques to obtain multi-source standard data of the region to be interpreted.

4. The natural ecological element classification method based on an adaptive deep forest model according to claim 3, characterized in that, The S3 natural element classification system consists of eight elements: mountains, water, forests, fields, lakes, grasslands, sand, and ice. Specifically, it includes: mountain element classification based on topographic relief areas; water element classification including linear and areal water bodies; forest element classification including natural and planted forests; field element classification including agricultural land and multi-cropping plots; lake element classification for independent water bodies; grassland element classification including natural grasslands, pastures, and low-lying grass and shrub vegetation; sand element classification; and ice element classification.

5. The natural ecological element classification method based on an adaptive deep forest model according to claim 1, characterized in that, S5 specifically includes: S51 Measurement Steps: Construct a local reference surface, determine the regional relative undulation of multi-source standard data and measure the uplift, and determine the relative undulation value; S52 discrimination steps: By using slope constraints and relative undulation thresholds, mountain discrimination is performed on multi-source standard data to determine the threshold discrimination rules; S53 Cleaning Steps: Morphology is introduced, and threshold discrimination rules are used to perform minimum spot cleaning on multi-source standard data to obtain the mountain extraction results.

6. The natural ecological element classification method based on an adaptive deep forest model according to claim 5, characterized in that, The formula for calculating the datum plane in S51 is: ； The formula for calculating relative fluctuations is: ； In the formula, Indicates the original DEM elevation. y Represents a pixel. Represented by pixels x The central area p Indicates the lower quantile. Represents a cell x The original DEM elevation, Represents a cell x The elevation of the reference surface.

7. The method for classifying natural ecological elements based on an adaptive deep forest model according to claim 5, characterized in that, The formula for calculating the slope constraint in S52 is: ； The formula for calculating the threshold discrimination rule is: ； In the formula, Indicates the original DEM elevation. x , y These represent the current cell and the cells in its neighborhood, respectively. express Preliminary judgment results R(x) Indicates relative fluctuation value, Indicates the relative fluctuation threshold. This is the slope threshold.

8. The natural ecological element classification method based on an adaptive deep forest model according to claim 1, characterized in that, The training process of the adaptive deep forest model in S6 specifically includes: Acquisition steps: Obtain the classification feature data for training and divide it into training set and validation set; Scanning steps: Perform multi-granularity feature scanning on the training set to obtain multi-dimensional feature vectors; Input steps: Input the multi-dimensional feature vector into the first layer of the cascaded forest structure random forest model for feature concatenation to obtain the first prediction vector; Training steps: For any layer after the first layer of the random forest model, the prediction vector of the previous layer is used as input to perform feature concatenation to obtain the prediction vector of the current layer, and the training result of the current layer is determined based on the validation set. Verification steps: Compare the training results of the current layer with the training results of the previous layer to obtain the performance improvement results; if the performance improvement results meet the preset threshold, then determine the current layer random forest model as the last layer random forest model and execute the next step; if the performance improvement results do not meet the preset threshold, then repeat the training steps. The steps are as follows: Based on the sparsity screening of logistic regression, invalid models from the first layer to the last layer of the random forest model are removed by Elastic Net regularization to obtain a trained adaptive deep forest model.

9. A method for classifying natural ecological elements based on an adaptive deep forest model according to claim 8, characterized in that, The specific formula for calculating the objective function of Elastic Net regularization is as follows: ； in, This represents the L1 norm, used to generate sparse solutions; λ represents the L2 norm, used to generate stable solutions; λ is the regularization strength, and λ≥0; ∈[0,1], used to control the ratio of L1 and L2.

10. A classification system for natural ecological elements based on an adaptive deep forest model, characterized in that, A method for classifying natural ecological elements based on an adaptive deep forest model as described in any one of claims 1-9 includes, in sequence, a data acquisition module, a processing module, a system setting module, a feature construction module, a mountain extraction module, a model classification module, and an overlay output module. The acquisition module is used to acquire classification feature data and multi-source remote sensing data of the area to be interpreted; The processing module is used to preprocess multi-source remote sensing data using affine transformation and resampling correlation techniques to obtain multi-source standard data; The system setting module is used to construct a classification system for natural elements based on traditional land use classification. The feature construction module is used to construct a classification feature dataset based on the classification system of natural elements and using classification element data. The classification feature dataset does not include mountain feature features; The mountain extraction module is used to construct a local reference surface, and to perform mountain discrimination and minimum patch cleaning on multi-source standard data using slope constraints and morphology to obtain mountain extraction results. The model classification module is used to input the classification feature dataset into the trained adaptive deep forest model to classify and interpret natural elements, and obtain a natural element classification result layer. The overlay output module is used to overlay the natural element classification result layer and the mountain extraction result through spatial analysis to obtain the final natural element classification result of the area to be interpreted.

Images