A Smart Classification Method and System for Complex Mountainous Topography Based on Dual-Scale TPI Optimization

Abstract

This invention relates to the field of geomorphological classification technology, and discloses an intelligent classification method and system for complex mountainous geomorphologies based on dual-scale TPI optimization. The method includes: first, acquiring ASTER GDEM data of the study area, dividing the terrain into analysis units and extracting initial terrain factors (large / small-scale terrain location indices, slope, and elevation); after standardization, calculating the Pearson correlation coefficient and variance inflation factor to diagnose factor collinearity; then, determining the optimal combination of terrain factors through screening and calibration; subsequently, matching this combination with historical geomorphological classification data of the study area to determine whether to adjust the input parameters of the machine learning model. If adjustment is needed, determining the primary and target influence coefficients of the model, and then optimizing the hyperparameters of random forest, extreme gradient boosting, and deep neural networks accordingly; finally, using the optimized model to complete the geomorphological classification of the study area and output the results. This invention improves the accuracy and efficiency of geomorphological classification in the study area by optimizing the combination of terrain factors and machine learning model parameters.

Description

Technical Field

[0001] This invention relates to the field of geomorphological classification technology, and more specifically, to an intelligent classification method and system for complex mountainous geomorphology based on dual-scale TPI optimization. Background Technology

[0002] Geomorphological classification is a key technology for dividing the Earth's surface into different geomorphic units based on morphological features. It is of great significance for revealing the interaction mechanism between topography and surface phenomena, supporting disaster prevention and control, ecological protection, and land use planning. Traditional geomorphological classification relies heavily on visual interpretation of topographic maps and aerial imagery, which requires field surveys. This not only consumes a lot of manpower and resources but also depends heavily on expert experience, limiting classification efficiency and accuracy.

[0003] With the development of Geographic Information Systems (GIS) and remote sensing technology, the application of Digital Elevation Models (DEMs) and remote sensing imagery has driven the development of automatic landform classification methods. However, existing methods still have shortcomings: single-scale Topographic Location Index (TPI) is difficult to take into account both macroscopic and microscopic landform features, resulting in low recognition accuracy for complex mountainous landforms; the selection of topographic factors lacks a systematic collinearity diagnosis mechanism, which easily reduces model efficiency due to factor redundancy; and the parameter optimization and classification result verification of machine learning models lack standardized procedures, making it difficult to reuse classification methods for different study areas.

[0004] Therefore, it is necessary to design a complex mountain landform intelligent classification method and system based on dual-scale TPI optimization to solve the problems of poor scale adaptability, unscientific factor selection, lack of standard model optimization, and lack of dynamic feedback mechanism in traditional landform classification. Summary of the Invention

[0005] In view of this, the present invention proposes an intelligent classification method and system for complex mountain landforms based on dual-scale TPI optimization, aiming to solve the problems of poor scale adaptability, unscientific factor selection, lack of standard model optimization, and lack of dynamic feedback mechanism in traditional landform classification.

[0006] In one aspect, this invention proposes an intelligent classification method for complex mountainous terrain based on dual-scale TPI optimization, comprising:

[0007] Acquire ASTER GDEM data of the study area, divide the data into multiple terrain analysis units, extract the initial terrain factors of each unit, standardize the initial terrain factors to obtain the characteristic data of each factor, calculate the Pearson correlation coefficient and variance inflation factor between the characteristic data, determine the degree of factor collinearity based on the Pearson correlation coefficient and variance inflation factor, and generate a collinearity diagnostic series.

[0008] Factors in the collinearity diagnostic sequence are screened and calibrated. The screening and calibration includes high collinearity calibration and low collinearity calibration. The number of factors with high collinearity calibration is counted, the number of factors with low collinearity calibration is counted, and the optimal combination of terrain factors is determined based on the proportion of the two types of calibration.

[0009] Historical geomorphological classification data of the study area is obtained, and the historical data is matched with the optimal factor combination. Based on the matching results, it is determined whether to adjust the input parameters of the machine learning model. When it is determined that adjustment is needed, the primary influence coefficient of the model is determined according to the weight ratio of the terrain factor, and the target influence coefficient of the model is determined in combination with the number of sample points.

[0010] Based on the model objective influence coefficient, determine the hyperparameter adjustment factors for random forest, extreme gradient boosting, and deep neural network. Optimize the model parameters based on these adjustment factors, and use the optimized model to classify the landforms of the study area, outputting the classification results.

[0011] Furthermore, after classifying the landforms of the study area using the optimized model and outputting the classification results, the process also includes:

[0012] Obtain spatial distribution data of various landform types in the classification results, extract boundary information and area proportion of each landform type based on the GIS system, and determine the accuracy verification index of landform classification based on the boundary information and area proportion;

[0013] A first precision threshold and a second precision threshold are preset, wherein the first precision threshold is less than the second precision threshold;

[0014] The accuracy verification index is compared with each preset accuracy threshold:

[0015] When the accuracy verification index is greater than or equal to the second accuracy threshold, the classification result is deemed qualified and the final landform classification map is directly output.

[0016] When the accuracy verification index is greater than or equal to the first accuracy threshold and less than the second accuracy threshold, local correction is performed on landform types with an area ratio of less than 5% in the classification results, and the classification map is output after correction.

[0017] When the accuracy verification index is less than the first accuracy threshold, the optimal combination of terrain factors is retrieved again, the hyperparameters of the machine learning model are adjusted, and the classification is performed again until the accuracy verification index is greater than or equal to the first accuracy threshold.

[0018] At the same time, the final classification results are overlaid with the historical geomorphological map of the study area to calculate the spatial fit. When the fit is greater than or equal to 85%, the classification process is completed.

[0019] When the agreement is less than 85%, the classification results are calibrated by combining field survey data.

[0020] Furthermore, when standardizing the initial terrain factors and generating the collinearity diagnostic series, the following steps are included:

[0021] Based on the collinearity diagnostic sequence, the variance inflation factor and tolerance of each initial terrain factor are extracted, and a collinearity determination threshold is set.

[0022] Compare the variance inflation factor and tolerance of each factor with the decision threshold:

[0023] A factor is marked as a highly collinear factor when it simultaneously satisfies the conditions that its variance inflation factor is greater than or equal to 10 and its tolerance is less than 0.1.

[0024] When a factor does not meet the requirements of variance inflation factor greater than or equal to 10 or tolerance greater than or equal to 0.1, it is marked as a low collinearity factor.

[0025] The proportion of highly collinear factors and lowly collinear factors is statistically analyzed. When the proportion of lowly collinear factors is greater than or equal to 70%, the set of lowly collinear factors is directly used as the optimal combination of terrain factors.

[0026] When the proportion of low collinearity factors is less than 70%, the collinearity of the remaining factors is recalculated after removing high collinearity factors until the proportion of low collinearity factors is ≥70%, thus forming the optimal combination of terrain factors.

[0027] Furthermore, determining the optimal combination of terrain factors includes:

[0028] A first factor quantity threshold and a second factor quantity threshold are preset, with the first factor quantity threshold being less than the second factor quantity threshold. The effectiveness of the combination is determined based on the relationship between the number of factors in the optimal terrain factor combination and the thresholds.

[0029] When the number of factors is greater than or equal to the threshold of the number of second factors, the combination is determined to contain sufficient terrain information and can be directly used for model training.

[0030] When the number of factors is greater than or equal to the first factor number threshold and less than the second factor number threshold, auxiliary factors such as surface cutting depth and slope variability are extracted until the number of factors is greater than or equal to the second factor number threshold.

[0031] When the number of factors is less than the first factor number threshold, the initial terrain factors are re-screened to expand the factor extraction range until the number of factors is greater than or equal to the first factor number threshold.

[0032] The final optimal combination of factors should include large-scale topographic location index, small-scale topographic location index, slope, elevation, surface cutting depth, slope variability, plan curvature, and profile curvature.

[0033] Furthermore, when determining the optimal analysis window for dual-scale TPI using the mean change point method, the following steps are included:

[0034] Based on the annular window, the initial inner and outer ring radii are set, and the window is gradually expanded at certain intervals. The mean and standard deviation of the terrain location index under each window are calculated.

[0035] Logarithmically transform the mean, construct a sample sequence, calculate the statistic using the mean change point method, and plot the curve of the difference in the statistic.

[0036] When the curve reaches an inflection point where it transitions from steep to gentle, the corresponding window is the optimal analysis window.

[0037] The optimal window for the small-scale topographic location index is the radius of the ring at the inflection point, corresponding to a certain area.

[0038] The optimal window for the large-scale topographic location index is the radius of a ring that has been expanded at certain intervals, corresponding to a certain area.

[0039] Furthermore, when matching historical geomorphological classification data with the optimal factor combination, the following steps are included:

[0040] Calculate the feature matching degree between historical classification data and the optimal factor combination, and preset the standard matching degree;

[0041] When the feature matching degree is consistent with the standard matching degree, the initial model parameters are used directly for training.

[0042] When the feature matching degree is inconsistent with the standard matching degree, the factor weights are adjusted according to the difference between the feature matching degree and the standard matching degree:

[0043] A first difference threshold and a second difference threshold are preset, and the first difference threshold is less than the second difference threshold;

[0044] When the difference between the feature matching degree and the standard matching degree is less than or equal to the first difference threshold, the weights of the large-scale terrain location index and the small-scale terrain location index are increased by 10%.

[0045] When the difference between the feature matching degree and the standard matching degree is greater than the first difference threshold and less than or equal to the second difference threshold, the weights of the surface cutting depth and slope are increased by 15%.

[0046] When the difference between the feature matching degree and the standard matching degree is greater than the second difference threshold, all factor weights are standardized and redistributed to ensure that the cumulative weight sum is 1.

[0047] Furthermore, when using the optimized model for landform classification, the following steps are included:

[0048] The sample points in the study area were divided into training and validation sets in a 7:3 ratio, and stratified random sampling was used to ensure that the proportion of various landform samples was consistent with the actual situation.

[0049] The classification results of the three models were evaluated using accuracy, area under the receiver operating characteristic (ROC) curve, recall, F1 score, and Kappa coefficient.

[0050] When the area under the receiver operating characteristic curve of the random forest is greater than or equal to 0.93 and the accuracy is greater than or equal to 87%, it is determined to be the optimal model, and its classification result is output.

[0051] When the random forest does not meet the above conditions but the area under the receiver operating characteristic curve of the deep neural network is greater than or equal to 0.90, the result of the deep neural network is adopted.

[0052] If neither of the two models satisfies the condition, the XGBoost iteration count is optimized and the model is reclassified until the area under the receiver operating characteristic curve is greater than or equal to 0.87.

[0053] Furthermore, when performing feature importance analysis on the classification results, the following is included:

[0054] The contribution of each factor to the optimal model is calculated based on the SHAP method, with a first contribution threshold and a second contribution threshold preset, and the first contribution threshold being greater than the second contribution threshold.

[0055] When the contribution of a factor is greater than or equal to the first contribution threshold, it is determined to be a key factor. The key factors include large-scale topographic location index, small-scale topographic location index, surface cutting depth and slope.

[0056] When the factor contribution is greater than or equal to the second contribution threshold and less than the first contribution threshold, it is determined to be an important factor. The important factors include elevation, plane curvature and profile curvature.

[0057] When the contribution of a factor is less than the second contribution threshold, it is determined to be a minor factor, and the minor factor includes the slope variability.

[0058] Among the key factors, the large-scale topographic location index contributes more than 50% to deeply incised canyons and U-shaped valleys, while the slope contributes more than 60% to plains.

[0059] Furthermore, when outputting the final landform classification results, the following are included:

[0060] The classification results were divided into ten landform types, including four macro-landform types and six micro-landform types.

[0061] Statistical analysis of the spatial distribution of various landforms:

[0062] Open sloping land accounts for the largest proportion of the area, mainly distributed in Chuxiong and central Kunming;

[0063] Deeply dissected canyons, U-shaped valleys, and other alpine-canyon landforms are concentrated in western Yuxi and northwestern Chuxiong.

[0064] The final output classification map should include heatmaps showing the boundaries, area proportions, and contribution of key factors for each landform type.

[0065] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0066] This application takes the central Yunnan region as the study area. Based on ASTER GDEM data, it extracts the dual-scale topographic location index (TPI) and various topographic factors, and combines machine learning models such as random forest, extreme gradient boosting and deep neural networks to carry out a complex landform classification study. Finally, random forest was determined to be the optimal model, with an overall accuracy of 87% and an AUC value of 0.93.

[0067] The optimal analysis window for large and small scales was determined by the mean change point method. The optimal combination of topographic factors was screened by combining Pearson correlation analysis and variance inflation factor collinearity diagnosis. The SHAP method was used to identify large-scale, small-scale, surface cutting depth and slope as key influencing factors.

[0068] This application effectively addresses the limitations of traditional single-scale TPI in the classification of complex landforms, and achieves the collaborative identification of macroscopic and microscopic landform features, providing scientific data support and methodological reference for disaster prevention, ecological protection and land use planning in central Yunnan.

[0069] On the other hand, this application also provides an intelligent classification system for complex mountain landforms based on dual-scale TPI optimization, applied to the aforementioned intelligent classification method for complex mountain landforms based on dual-scale TPI optimization, including:

[0070] The data preprocessing module is used to acquire ASTER GDEM data of the study area, divide the data into multiple terrain analysis units, extract the initial terrain factors of each unit, the terrain includes large-scale terrain location index, small-scale terrain location index, slope, and elevation, standardize the initial terrain factors to obtain the feature data of each factor, calculate the Pearson correlation coefficient and variance inflation factor between the feature data, determine the factor collinearity based on the Pearson correlation coefficient and variance inflation factor, and generate a collinearity diagnostic series.

[0071] The factor screening module is used to screen and calibrate the factors in the collinearity diagnostic sequence. The screening and calibration includes high collinearity calibration and low collinearity calibration. The number of factors with high collinearity calibration is counted, the number of factors with low collinearity calibration is counted, and the optimal combination of terrain factors is determined based on the proportion of the two types of calibration.

[0072] The model parameter tuning module is used to acquire historical geomorphological classification data of the study area, match the historical data with the optimal factor combination, and determine whether to adjust the input parameters of the machine learning model based on the matching results. When it is determined that adjustment is needed, the primary influence coefficient of the model is determined based on the weight ratio of the terrain factor, and the target influence coefficient of the model is determined in combination with the number of sample points.

[0073] The classification output module is used to determine the hyperparameter adjustment factors of random forest, extreme gradient boosting and deep neural network according to the model target influence coefficient, optimize the model parameters according to the adjustment factors, use the optimized model to classify the landform of the study area, and output the classification results.

[0074] It is understandable that the aforementioned intelligent classification method and system for complex mountainous terrain based on dual-scale TPI optimization has the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0075] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0076] Figure 1 A flowchart illustrating the intelligent classification method for complex mountainous terrain based on dual-scale TPI optimization provided in this embodiment of the invention;

[0077] Figure 2 This is a functional block diagram of a complex mountainous terrain intelligent classification system based on dual-scale TPI optimization provided in an embodiment of the present invention. Detailed Implementation

[0078] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0079] Reference Figure 1 In some embodiments of this application, a method for intelligent classification of complex mountainous terrain based on dual-scale TPI optimization includes:

[0080] ASTER GDEM data of the study area was acquired, the data was divided into multiple terrain analysis units, the initial terrain factors of each unit were extracted, the initial terrain factors were standardized to obtain the characteristic data of each factor, the Pearson correlation coefficient and variance inflation factor between the characteristic data were calculated, the degree of factor collinearity was determined based on the Pearson correlation coefficient and variance inflation factor, and a collinearity diagnostic series was generated.

[0081] Factors in the collinearity diagnostic series are screened and calibrated. Screening and calibration include high collinearity calibration and low collinearity calibration. The number of factors with high collinearity calibration is counted, and the number of factors with low collinearity calibration is counted. The optimal combination of terrain factors is determined based on the proportion of the two types of calibration.

[0082] Historical geomorphological classification data of the study area is obtained, and the historical data is matched with the optimal factor combination. Based on the matching results, it is determined whether to adjust the input parameters of the machine learning model. When it is determined that adjustment is needed, the primary influence coefficient of the model is determined according to the weight ratio of the topographic factors, and the target influence coefficient of the model is determined in combination with the number of sample points.

[0083] Based on the model objective influence coefficient, determine the hyperparameter adjustment factors for random forest, extreme gradient boosting, and deep neural network. Optimize the model parameters based on the adjustment factors, and use the optimized model to classify the landforms of the study area, outputting the classification results.

[0084] Specifically, to comprehensively consider both macroscopic and microscopic features of landforms, and referencing previous research, a total of 11 topographic factors were selected. These include six macroscopic factors at two scales: Topographic Location Index (TPI), Slope (S), Elevation (E), Plane Curvature (PLC), Profile Curvature (PC), and Topographic Relief (RF); and four microscopic topographic factors: Surface Cut Depth (SCD), Surface Roughness (SR), Elevation Variation Coefficient (EVC), and Slope Variation Rate (SOS). The calculation formulas for each factor are shown in Equations (1-5). To reduce data redundancy and improve model computational efficiency and classification accuracy, this study conducted correlation analysis and collinearity diagnosis on the topographic factors derived from DEM data to identify and eliminate highly correlated variables.

[0085]

[0086] In equation (1): E std E mean E min E maxThese represent the standard deviation, average, maximum, and minimum elevation values ​​within the statistical window, respectively.

[0087]

[0088] In equation (2): ΔZ, ΔD, and ΔS represent the elevation difference (i.e., height difference) between adjacent pixels, the horizontal distance (i.e., horizontal projection distance) between adjacent pixels, and the slope difference (i.e., the change in slope) between adjacent pixels, respectively.

[0089]

[0090] In equation (3): Z and s represent the elevation value, the distance along the slope direction of the curve, and x and y are the coordinates in the horizontal direction, respectively.

[0091] RF=Z max -Z min (4)

[0092] In equation (4): Z max Z min These represent the highest and lowest elevation values ​​within the region, respectively.

[0093] Specifically, the Topographic Position Index (TPI) is defined by calculating the difference between the elevation of a cell and the average elevation of cells within a predetermined radius around it. A positive TPI value indicates high ground (ridge), a negative value indicates low ground (valley), and zero indicates a flat area. TPI can effectively distinguish the slope of terrain (such as ridges, valleys, and hillsides) and, when combined with slope data, can further identify terrain morphology types (such as canyons, plains, and U-shaped valleys). The extraction formula is:

[0094]

[0095] In equation (5): Z cell N, Z i These represent the elevation value of the center cell, the number of cells in the neighboring area, and the elevation values ​​of other cells in the neighboring area, respectively.

[0096] The above embodiments take the central Yunnan region as the study area. Based on ASTER GDEM data, a complex landform classification study was conducted by extracting the dual-scale Topographic Location Index (TPI) and various topographic factors, combined with machine learning models such as random forest, extreme gradient boosting, and deep neural networks. Random forest was ultimately determined as the optimal model, achieving an overall accuracy of 87% and an AUC of 0.93. The optimal analysis windows at large and small scales were determined using the mean-variable method. The optimal combination of topographic factors was screened using Pearson correlation analysis and variance inflation factor collinearity diagnosis. The SHAP method was used to identify large-scale, small-scale, surface incision depth, and slope as key influencing factors. This application effectively overcomes the limitations of traditional single-scale TPI in complex landform classification, achieving collaborative identification of macroscopic and microscopic landform features, and providing scientific data support and methodological reference for disaster prevention, ecological protection, and land use planning in the central Yunnan region.

[0097] Specifically, after classifying the landforms of the study area using the optimized model and outputting the classification results, the process also includes:

[0098] Obtain spatial distribution data of various landform types in the classification results, extract boundary information and area proportion of each landform type based on the GIS system, and determine the accuracy verification index of landform classification based on the boundary information and area proportion;

[0099] A first precision threshold and a second precision threshold are preset, wherein the first precision threshold is less than the second precision threshold;

[0100] The accuracy verification index is compared with each preset accuracy threshold:

[0101] When the accuracy verification index is greater than or equal to the second accuracy threshold, the classification result is deemed qualified and the final landform classification map is directly output.

[0102] When the accuracy verification index is greater than or equal to the first accuracy threshold and less than the second accuracy threshold, local correction is performed on landform types with an area ratio of less than 5% in the classification results, and the classification map is output after correction.

[0103] When the accuracy verification index is less than the first accuracy threshold, the optimal combination of terrain factors is retrieved again, the hyperparameters of the machine learning model are adjusted, and the classification is performed again until the accuracy verification index is greater than or equal to the first accuracy threshold.

[0104] At the same time, the final classification results are overlaid with the historical geomorphological map of the study area to calculate the spatial fit. When the fit is greater than or equal to 85%, the classification process is completed.

[0105] When the agreement is less than 85%, the classification results are calibrated by combining field survey data.

[0106] It is understandable that the above embodiments take the central Yunnan region as the study area. Based on ASTER GDEM data, dual-scale topographic location index (TPI) and various topographic factors are extracted. The optimal factor combination is screened through Pearson correlation analysis and variance inflation factor collinearity diagnosis. Random forest, extreme gradient boosting and deep neural network are used for landform classification. Random forest is determined to be the optimal model (overall accuracy 87%, AUC 0.93). The SHAP method is used to identify large-scale TPI, small-scale TPI, surface incision depth and slope as key factors. Finally, a classification result that takes into account both macro and micro landform characteristics is formed, providing data support for disaster prevention and control in this region.

[0107] Specifically, the process of standardizing the initial terrain factors and generating a collinearity diagnostic series includes:

[0108] Based on the collinearity diagnostic sequence, the variance inflation factor and tolerance of each initial terrain factor are extracted, and a collinearity judgment threshold is set.

[0109] Compare the variance inflation factor and tolerance of each factor with the decision threshold:

[0110] A factor is marked as a highly collinear factor when it simultaneously satisfies the conditions that its variance inflation factor is greater than or equal to 10 and its tolerance is less than 0.1.

[0111] When a factor does not meet the requirements of variance inflation factor greater than or equal to 10 or tolerance greater than or equal to 0.1, it is marked as a low collinearity factor.

[0112] The proportion of highly collinear factors and lowly collinear factors is statistically analyzed. When the proportion of lowly collinear factors is greater than or equal to 70%, the set of lowly collinear factors is directly used as the optimal combination of terrain factors.

[0113] When the proportion of low collinearity factors is less than 70%, the collinearity of the remaining factors is recalculated after removing high collinearity factors until the proportion of low collinearity factors is ≥70%, thus forming the optimal combination of terrain factors.

[0114] Specifically, when standardizing the initial topographic factors and generating a collinearity diagnostic series, the variance inflation factor and tolerance of each factor are extracted. The collinearity judgment threshold is set as variance inflation factor ≥ 10 and tolerance < 0.1. Factors that meet this condition are high collinearity factors, and those that do not are low collinearity factors. When the proportion of low collinearity factors is ≥ 70%, it is directly used as the optimal combination. When the proportion is < 70%, high collinearity factors are removed and collinearity is recalculated until the proportion of low collinearity factors is ≥ 70%. The final optimal factor combination includes topographic location index (large and small scales), surface cutting depth, slope, elevation, slope variability, plane curvature, and profile curvature.

[0115] The above embodiments, by combining the optimal dual-scale TPI with machine learning models, have achieved accurate classification of complex landforms in central Yunnan. This approach takes into account both macroscopic and microscopic landform characteristics, overcoming the shortcomings of traditional single-scale methods. Furthermore, it improves classification accuracy (overall accuracy of 87% for random forest and AUC 0.93) through factor screening and model optimization, identifies key influencing factors, and provides reliable data support and scientific basis for regional disaster prevention, ecological protection, and land use planning.

[0116] Specifically, determining the optimal combination of terrain factors includes:

[0117] A first factor quantity threshold and a second factor quantity threshold are preset, with the first factor quantity threshold being less than the second factor quantity threshold. The effectiveness of the combination is determined based on the relationship between the number of factors in the optimal terrain factor combination and the thresholds.

[0118] When the number of factors is greater than or equal to the threshold of the number of second factors, the combination is determined to contain sufficient terrain information and can be directly used for model training.

[0119] When the number of factors is greater than or equal to the first factor number threshold and less than the second factor number threshold, auxiliary factors such as surface cutting depth and slope variability are extracted until the number of factors is greater than or equal to the second factor number threshold.

[0120] When the number of factors is less than the first factor number threshold, the initial terrain factors are re-screened to expand the factor extraction range until the number of factors is greater than or equal to the first factor number threshold.

[0121] The final optimal combination of factors should include large-scale topographic location index, small-scale topographic location index, slope, elevation, surface cutting depth, slope variability, plan curvature, and profile curvature.

[0122] Specifically, when determining the optimal combination of topographic factors, a threshold of 5 is preset for the number of the first factor and 8 is preset for the number of the second factor (the first threshold < the second threshold). When the number of factors is ≥8, they are directly used for model training. When 5 ≤ number of factors < 8, auxiliary factors such as surface cutting depth and slope variability are extracted to bring the number of factors to ≥8. When the number of factors is <5, the initial topographic factors are re-screened and the extraction range is expanded to bring the number of factors to ≥5. The final optimal combination of factors includes a large-scale topographic location index, a small-scale topographic location index, slope, elevation, surface cutting depth, slope variability, plan curvature, and profile curvature, for a total of 8 factors.

[0123] The above embodiments determine the optimal large- and small-scale TPI analysis windows through the mean-variable method, and select the optimal combination of topographic factors including dual-scale TPI, slope, and elevation by combining Pearson correlation and variance inflation factor. Classification is achieved using models such as random forest, with the overall accuracy of random forest reaching 87% and an AUC value of 0.93. It also identifies dual-scale TPI, surface incision depth, and slope as key factors, effectively taking into account both macro and micro geomorphic features. This solves the shortcomings of traditional single-scale methods in the classification of complex geomorphologies and provides accurate geomorphic data support for disaster prevention, ecological protection, and land use planning in this region.

[0124] Specifically, when determining the optimal analysis window for dual-scale TPI using the mean change point method, the following steps are included:

[0125] Based on the annular window, the initial inner and outer ring radii are set, and the window is gradually expanded at certain intervals. The mean and standard deviation of the terrain location index under each window are calculated.

[0126] Logarithmically transform the mean, construct a sample sequence, calculate the statistic using the mean change point method, and plot the curve of the difference in the statistic.

[0127] When the curve reaches an inflection point where it transitions from steep to gentle, the corresponding window is the optimal analysis window.

[0128] The optimal window for the small-scale topographic location index is the radius of the ring at the inflection point, corresponding to a certain area.

[0129] The optimal window for the large-scale topographic location index is the radius of a ring that has been expanded at certain intervals, corresponding to a certain area.

[0130] Specifically, the TPI analysis window includes circular, annular, fan-shaped, and rectangular shapes. Due to the complex terrain features in central Yunnan, the annular window can better capture such radial or curved terrain structures, and is particularly adaptable to the analysis of terrain types such as ridges, valleys, and slopes. Therefore, this study chose the annular window for analysis.

[0131] Specifically, within a specific geomorphic region, to determine the optimal TPI value for a given point, this study uses a ring-shaped window as a basis, gradually expanding the radius for analysis. As the analysis area increases, the TPI gradually increases at surface protrusions and gradually decreases at depressions. When the analysis area reaches a certain threshold, the change in the neighborhood's average elevation tends to stabilize. The inflection point where the rate of increase in average elevation slows down is defined, and its corresponding area is the statistical unit. Methods for interpreting inflection points include the standard deviation method, the fitted curve method, and the change-point method. Among these, the mean change-point analysis method is widely used because it can scientifically and concisely calculate the location of the inflection point on the fitted curve where the steepness gradually decreases. The average elevation within this statistical unit can reasonably characterize the relative topographical position of the research point.

[0132] The TPI analysis window includes circular, annular, fan-shaped, and rectangular shapes. Due to the complex terrain features in central Yunnan, the annular window can better capture such radial or curved terrain structures, and is particularly adaptable to the analysis of terrain types such as ridges, valleys, and slopes. Therefore, this study chose the annular window for analysis.

[0133] In a specific geomorphic region, in order to determine the optimal TPI value of a certain point, this study uses an annular window as the basis and gradually expands the radius for analysis. As the analysis area increases, the TPI of the surface protrusions gradually increases, while that of the depressions gradually decreases. When the analysis area increases to a certain threshold, the change in the average elevation of the neighborhood tends to be stable. The inflection point where the growth rate of the average elevation changes from fast to slow is defined, and the corresponding area is the statistical unit. The methods for interpreting the inflection point include the standard deviation method, the fitting curve method, and the change point method. Among them, the mean change point analysis method

[42] is widely used because it can scientifically and concisely calculate the position of the inflection point on the fitting curve from steep to slow. The average elevation within this statistical unit can reasonably characterize the relative topographic position of the research point.

[0134] An initial analysis window with an inner ring radius of 1 and an outer ring radius of 3 was used. The inner and outer ring radii were gradually expanded at intervals of 2 pixels until the inner ring radius reached 55 and the outer ring radius reached 57. For analysis windows of different scales, the average value of the topographic location index (TPIm), the area of ​​the statistical unit, and the standard deviation (SD) were statistically analyzed (Table 1).

[0135] Table 1. Statistical characteristics of TPI under different analysis windows

[0136]

[0137]

[0138] This study uses the toolbox module of ArcMap 10.8 and the Focal Statistics and Neighborhood tools to process DEM data from the central Yunnan region and extract TPI data. A circular analysis window is selected, centered on the target raster. Starting with an inner and outer ring radius of (1,3), the inner and outer ring radii are gradually increased until the inner and outer ring radii reach (55,57).

[0139] Since the above analysis of inflection points is based on human judgment and may be influenced by subjective factors, the results contain a certain degree of error. The logarithm of the TPI mean in Table 1 is used to construct the sample sequence {x}. i}, where i = 1, 2, ..., N, and N is the sample size. The samples are defined by x. i Divide the data into two segments using the point as the boundary, and calculate the statistic using the following formula:

[0140]

[0141] In equations (6) and (7), and Let each represent the arithmetic mean of the sample. S represents the average of the entire sample. i S and S represent the intermediate calculated values ​​of the mean-changing point method, respectively.

[0142] 2.2.3 Construction of a Classification System for Landform Types

[0143] After determining the optimal TPI analysis window, this window is selected as the small-scale TPI (FTPI). Starting from this window, the inner and outer ring radii are gradually expanded at intervals of 5 pixels. The large-scale TPI (BTPI) is determined using the same method. Combining the small-scale TPI with the large-scale TPI helps to distinguish different types of nested terrain features. Referring to Weiss's research

[32] , the obtained FTPI and BTPI are standardized according to equation (8):

[0144]

[0145] In equation (6): μ and σ represent the mean and standard deviation of TPI, respectively. That is, standardize the TPI so that its mean is 0 and its standard deviation is 1. ×100 means scaling the standardized value to the integer range, and scale +0.5 and int(...) mean rounding to the nearest integer. Referring to the classification schemes of previous studies

[43] , the landforms of central Yunnan are then classified according to 1 standard deviation unit (=100 grid value units) as shown in Table (2):

[0146] Table 2. Criteria for Classification of Landform Types

[0147]

[0148] Specifically, when matching historical geomorphological classification data with the optimal factor combination, the following is included:

[0149] Calculate the feature matching degree between historical classification data and the optimal factor combination, and preset the standard matching degree;

[0150] When the feature matching degree is consistent with the standard matching degree, the initial model parameters are used directly for training.

[0151] When the feature matching degree is inconsistent with the standard matching degree, the factor weights are adjusted according to the difference between the feature matching degree and the standard matching degree:

[0152] A first difference threshold and a second difference threshold are preset, and the first difference threshold is less than the second difference threshold;

[0153] When the difference between the feature matching degree and the standard matching degree is less than or equal to the first difference threshold, the weights of the large-scale terrain location index and the small-scale terrain location index are increased by 10%.

[0154] When the difference between the feature matching degree and the standard matching degree is greater than the first difference threshold and less than or equal to the second difference threshold, the weights of the surface cutting depth and slope are increased by 15%.

[0155] When the difference between the feature matching degree and the standard matching degree is greater than the second difference threshold, all factor weights are standardized and redistributed to ensure that the cumulative weight sum is 1.

[0156] The above embodiments calculate the feature matching degree between historical classification data and the optimal factor combination, and preset a standard matching degree. If the feature matching degree and the standard matching degree are consistent, it means that the current data and factor combination have a good fit, and the initial model parameters can be directly used for training to ensure that the model starts smoothly on the basis of existing data. When the two are inconsistent, the factor weights need to be adjusted according to the difference. A first difference threshold of 0.1 and a second difference threshold of 0.2 are preset (the first difference threshold is less than the second difference threshold). When the difference between the feature matching degree and the standard matching degree is less than or equal to 0.1, the weights of the large-scale topographic location index and the small-scale topographic location index are increased by 10% to enhance their influence in the model and better adapt to the data, since the large-scale topographic location index has obvious advantages in reflecting the comprehensiveness and uniqueness of topographic features. When the difference is greater than 0.1 and less than or equal to 0.2, the surface cutting depth and slope play a key role in the classification of topographic morphology and landform type. At this time, their weights are increased by 15% to make the model pay more attention to the role of these key topographic factors. When the difference is greater than 0.2, it indicates that the overall data deviation is large. All factor weights are standardized and redistributed, and mathematical methods are used to ensure that the cumulative weight sum is always 1, so that each factor can find a new balance in the model, optimize the model's fit to the data, and thus improve the accuracy of complex mountain landform classification.

[0157] Specifically, when using the optimized model for terrain classification, the following are included:

[0158] The sample points in the study area were divided into training and validation sets in a 7:3 ratio, and stratified random sampling was used to ensure that the proportion of various landform samples was consistent with the actual situation.

[0159] The classification results of the three models were evaluated using accuracy, area under the receiver operating characteristic (ROC) curve, recall, F1 score, and Kappa coefficient.

[0160] When the area under the receiver operating characteristic curve of the random forest is greater than or equal to 0.93 and the accuracy is greater than or equal to 87%, it is determined to be the optimal model, and its classification result is output.

[0161] When the random forest does not meet the above conditions but the area under the receiver operating characteristic curve of the deep neural network is greater than or equal to 0.90, the result of the deep neural network is adopted.

[0162] If neither of the two models satisfies the condition, the XGBoost iteration count is optimized and the model is reclassified until the area under the receiver operating characteristic curve is greater than or equal to 0.87.

[0163] Understandably, the Random Forest (RF) classifier achieves classification by integrating the predictions of multiple decision trees. Its advantage lies in its ability to effectively handle the nonlinear and complex interactions between variables in high-dimensional geomorphological data, thus it is considered a reliable and robust classification method. XGBoost is a machine learning algorithm based on gradient boosting trees. This algorithm has demonstrated excellent performance in classification and regression tasks, combining flexibility and high scalability, and has become one of the most influential and widely used algorithms in the field of machine learning. With the significant improvement in computing power and the continuous expansion of data scale, deep learning methods have developed rapidly. Among them, the application of deep neural networks (DNNs) in geomorphological classification can significantly improve classification accuracy and efficiency, especially when combined with high-resolution DEM data.

[0164] Specifically, when performing feature importance analysis on the classification results, the following is included:

[0165] The contribution of each factor to the optimal model is calculated based on the SHAP method, with a first contribution threshold and a second contribution threshold preset, and the first contribution threshold being greater than the second contribution threshold.

[0166] When the contribution of a factor is greater than or equal to the first contribution threshold, it is determined to be a key factor. Key factors include large-scale topographic location index, small-scale topographic location index, surface cutting depth and slope.

[0167] When the factor contribution is greater than or equal to the second contribution threshold and less than the first contribution threshold, it is determined to be an important factor. Important factors include elevation, plane curvature and profile curvature.

[0168] When the contribution of a factor is less than the second contribution threshold, it is determined to be a minor factor. Minor factors include slope variability.

[0169] Among the key factors, the large-scale topographic location index contributes more than 50% to deeply incised canyons and U-shaped valleys, while the slope contributes more than 60% to plains.

[0170] Understandably, the above embodiments use the SHAP method to perform feature importance analysis on the classification results, clarifying key factors (large-scale, small-scale, surface incision depth, slope), important factors, and minor factors, as well as their contribution to different landform types (e.g., large-scale TPI contributes over 50% to deeply incised canyons, and slope contributes over 60% to plains). This not only reveals the intrinsic mechanism by which topographic factors affect landform classification, but also provides a scientific basis for the selection and application of factors in the classification of complex mountain landforms, improving the interpretability and reliability of the classification results.

[0171] Specifically, when outputting the final landform classification results, the following are included:

[0172] The classification results were divided into ten landform types, including four macro-landform types and six micro-landform types.

[0173] Statistical analysis of the spatial distribution of various landforms:

[0174] Open sloping land accounts for the largest proportion of the area, mainly distributed in Chuxiong and central Kunming;

[0175] Deeply dissected canyons, U-shaped valleys, and other alpine-canyon landforms are concentrated in western Yuxi and northwestern Chuxiong.

[0176] The final output classification map should include heatmaps showing the boundaries, area proportions, and contribution of key factors for each landform type.

[0177] Understandably, this classification accurately divides the landforms of central Yunnan into 10 types, clarifying the spatial distribution characteristics of macro and micro landforms (such as open slopes mainly distributed in Chuxiong and central Kunming, and high mountain-canyon landforms concentrated in western Yuxi). Moreover, the output classification map includes heat maps of boundaries, area proportions, and contributions of key factors. This not only refines the spatial pattern of complex landforms but also provides intuitive and accurate geomorphological data support for regional disaster prevention, ecological protection, and land use planning, thereby enhancing the application value of the classification results.

[0178] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides an intelligent classification system for complex mountainous terrain based on dual-scale TPI optimization, including:

[0179] The data preprocessing module is used to acquire ASTER GDEM data of the study area, divide the data into multiple terrain analysis units, extract the initial terrain factors of each unit, including large-scale terrain location index, small-scale terrain location index, slope, and elevation, standardize the initial terrain factors to obtain the feature data of each factor, calculate the Pearson correlation coefficient and variance inflation factor between the feature data, determine the degree of factor collinearity based on the Pearson correlation coefficient and variance inflation factor, and generate a collinearity diagnostic series.

[0180] The factor screening module is used to screen and calibrate factors in the collinearity diagnostic series. The screening and calibration include high collinearity calibration and low collinearity calibration. The number of factors with high collinearity calibration is counted, the number of factors with low collinearity calibration is counted, and the optimal combination of terrain factors is determined based on the proportion of the two types of calibration.

[0181] The model parameter tuning module is used to acquire historical geomorphological classification data of the study area, match the historical data with the optimal factor combination, and determine whether to adjust the input parameters of the machine learning model based on the matching results. When it is determined that adjustment is needed, the primary influence coefficient of the model is determined based on the weight ratio of the terrain factor, and the target influence coefficient of the model is determined in combination with the number of sample points.

[0182] The classification output module is used to determine the hyperparameter adjustment factors of random forest, extreme gradient boosting and deep neural network based on the model objective influence coefficient, optimize the model parameters according to the adjustment factors, use the optimized model to classify the landform of the study area, and output the classification results.

[0183] It is understandable that the aforementioned intelligent classification method and system for complex mountainous terrain based on dual-scale TPI optimization has the same beneficial effects, and will not be elaborated further here.

[0184] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0185] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0186] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0187] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0188] 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 scope of protection of the claims of the present invention.

Claims

1. A method for intelligent classification of complex mountainous terrain based on dual-scale TPI optimization, characterized in that, include: Acquire ASTER GDEM data of the study area, divide the data into multiple terrain analysis units, and extract the initial terrain factors for each unit; Based on the annular window, the initial inner and outer ring radii are set, and the window is gradually expanded at certain intervals. The mean and standard deviation of the terrain location index under each window are calculated. Logarithmically transform the mean, construct a sample sequence, calculate the statistic using the mean change point method, and plot the curve of the difference in the statistic. When the curve reaches an inflection point where it becomes gentler, the corresponding window is the optimal analysis window for the small-scale terrain location index. The optimal small-scale window is the radius of the ring at the inflection point, corresponding to a certain area. When determining the optimal combination of terrain factors, the final optimal combination of factors must include both large-scale terrain location index and small-scale terrain location index. Starting from the small-scale optimal analysis window, the inner and outer ring radii are gradually expanded at intervals of 5 pixels. The same mean-variable point method is used to determine the optimal analysis window for the large-scale terrain location index. The large-scale optimal window is the ring radius at the corresponding inflection point, which corresponds to a certain area. The initial terrain factors are standardized to obtain the characteristic data of each factor. The Pearson correlation coefficient and variance inflation factor between the characteristic data are calculated. The degree of factor collinearity is determined based on the Pearson correlation coefficient and variance inflation factor, and a collinearity diagnostic series is generated. Factors in the collinearity diagnostic sequence are screened and calibrated. The screening and calibration includes high collinearity calibration and low collinearity calibration. The number of factors with high collinearity calibration is counted, the number of factors with low collinearity calibration is counted, and the optimal combination of terrain factors is determined based on the proportion of the two types of calibration. Historical geomorphological classification data of the study area is obtained, and the historical geomorphological classification data is matched with the optimal factor combination. Based on the matching results, it is determined whether to adjust the input parameters of the machine learning model. When it is determined that adjustment is needed, the primary influence coefficient of the model is determined according to the weight ratio of the terrain factor, and the target influence coefficient of the model is determined in combination with the number of sample points. Based on the model objective influence coefficient, determine the hyperparameter adjustment factors for random forest, extreme gradient boosting, and deep neural network. Optimize the model parameters based on these adjustment factors, and use the optimized model to classify the landforms of the study area, outputting the classification results.

2. The intelligent classification method for complex mountainous terrain based on dual-scale TPI optimization according to claim 1, characterized in that, After classifying the landforms of the study area using the optimized model and outputting the classification results, the process also includes: Obtain spatial distribution data of various landform types in the classification results, extract boundary information and area proportion of each landform type based on the GIS system, and determine the accuracy verification index of landform classification based on the boundary information and area proportion; A first precision threshold and a second precision threshold are preset, wherein the first precision threshold is less than the second precision threshold; The accuracy verification index is compared with each preset accuracy threshold: When the accuracy verification index is greater than or equal to the second accuracy threshold, the classification result is deemed qualified and the final landform classification map is directly output. When the accuracy verification index is greater than or equal to the first accuracy threshold and less than the second accuracy threshold, local correction is performed on landform types with an area ratio of less than 5% in the classification results, and the classification map is output after correction. When the accuracy verification index is less than the first accuracy threshold, the optimal combination of terrain factors is retrieved again, the hyperparameters of the machine learning model are adjusted, and the classification is performed again until the accuracy verification index is greater than or equal to the first accuracy threshold. At the same time, the final classification results are overlaid with the historical geomorphological map of the study area to calculate the spatial fit. When the fit is greater than or equal to 85%, the classification process is completed. When the agreement is less than 85%, the classification results are calibrated by combining field survey data.

3. The intelligent classification method for complex mountainous terrain based on dual-scale TPI optimization according to claim 2, characterized in that, The process of standardizing the initial terrain factors and generating the collinearity diagnostic series includes: Based on the collinearity diagnostic sequence, the variance inflation factor and tolerance of each initial terrain factor are extracted, and a collinearity determination threshold is set. Compare the variance inflation factor and tolerance of each factor with the decision threshold: A factor is marked as a highly collinear factor when it simultaneously satisfies the conditions that its variance inflation factor is greater than or equal to 10 and its tolerance is less than 0.

1. When a factor does not meet the requirements of variance inflation factor greater than or equal to 10 or tolerance greater than or equal to 0.1, it is marked as a low collinearity factor. The proportion of highly collinear factors and lowly collinear factors is statistically analyzed. When the proportion of lowly collinear factors is greater than or equal to 70%, the set of lowly collinear factors is directly used as the optimal combination of terrain factors. When the proportion of low collinearity factors is less than 70%, the collinearity of the remaining factors is recalculated after removing high collinearity factors until the proportion of low collinearity factors is ≥70%, thus forming the optimal combination of terrain factors.

4. The intelligent classification method for complex mountainous terrain based on dual-scale TPI optimization according to claim 3, characterized in that, Determining the optimal combination of terrain factors includes: A first factor quantity threshold and a second factor quantity threshold are preset, with the first factor quantity threshold being less than the second factor quantity threshold. The effectiveness of the combination is determined based on the relationship between the number of factors in the optimal terrain factor combination and the thresholds. When the number of factors is greater than or equal to the threshold of the number of second factors, the combination is determined to contain sufficient terrain information and can be directly used for model training. When the number of factors is greater than or equal to the first factor number threshold and less than the second factor number threshold, supplementary factors such as surface cutting depth and slope variability are extracted until the number of factors is greater than or equal to the second factor number threshold. When the number of factors is less than the first factor number threshold, the initial terrain factors are re-screened to expand the factor extraction range until the number of factors is greater than or equal to the first factor number threshold. The final optimal combination of factors should include large-scale topographic location index, small-scale topographic location index, slope, elevation, surface cutting depth, slope variability, plan curvature, and profile curvature.

5. The intelligent classification method for complex mountainous terrain based on dual-scale TPI optimization according to claim 4, characterized in that, When matching historical geomorphological classification data with the optimal factor combination, the following is included: Calculate the feature matching degree between historical classification data and the optimal factor combination, and preset the standard matching degree; When the feature matching degree is consistent with the standard matching degree, the initial model parameters are used directly for training. When the feature matching degree is inconsistent with the standard matching degree, the factor weights are adjusted according to the difference between the feature matching degree and the standard matching degree: A first difference threshold and a second difference threshold are preset, and the first difference threshold is less than the second difference threshold; When the difference between the feature matching degree and the standard matching degree is less than or equal to the first difference threshold, the weights of the large-scale terrain location index and the small-scale terrain location index are increased by 10%. When the difference between the feature matching degree and the standard matching degree is greater than the first difference threshold and less than or equal to the second difference threshold, the weights of the surface cutting depth and slope are increased by 15%. When the difference between the feature matching degree and the standard matching degree is greater than the second difference threshold, all factor weights are standardized and redistributed to ensure that the cumulative weight sum is 1.

6. The intelligent classification method for complex mountainous terrain based on dual-scale TPI optimization according to claim 5, characterized in that, When using the optimized model for terrain classification, the following are included: The sample points in the study area were divided into training and validation sets in a 7:3 ratio, and stratified random sampling was used to ensure that the proportion of various landform samples was consistent with the actual situation. The classification results of the three models were evaluated using accuracy, area under the receiver operating characteristic (ROC) curve, recall, F1 score, and Kappa coefficient. When the area under the receiver operating characteristic curve of the random forest is greater than or equal to 0.93 and the accuracy is greater than or equal to 87%, it is determined to be the optimal model and its classification result is output. When the random forest does not meet the conditions but the area under the receiver operating characteristic curve of the deep neural network is greater than or equal to 0.90, the result of the deep neural network is adopted. If neither of the two models satisfies the condition, the XGBoost iteration count is optimized and the model is reclassified until the area under the receiver operating characteristic curve is greater than or equal to 0.

87.

7. The intelligent classification method for complex mountainous terrain based on dual-scale TPI optimization according to claim 6, characterized in that, When performing feature importance analysis on the classification results, the following should be included: The contribution of each factor to the optimal model is calculated based on the SHAP method, with a first contribution threshold and a second contribution threshold preset, and the first contribution threshold being greater than the second contribution threshold. When the contribution of a factor is greater than or equal to the first contribution threshold, it is determined to be a key factor. The key factors include large-scale topographic location index, small-scale topographic location index, surface cutting depth and slope. When the factor contribution is greater than or equal to the second contribution threshold and less than the first contribution threshold, it is determined to be an important factor. The important factors include elevation, plane curvature and profile curvature. When the contribution of a factor is less than the second contribution threshold, it is determined to be a minor factor, and the minor factor includes the slope variability. Among the key factors, the large-scale topographic location index contributes more than 50% to deeply incised canyons and U-shaped valleys, while the slope contributes more than 60% to plains.

8. The intelligent classification method for complex mountainous terrain based on dual-scale TPI optimization according to claim 7, characterized in that, When outputting the final landform classification results, the following should be included: The classification results were divided into ten landform types, including four macro-landform types and six micro-landform types. Statistical analysis of the spatial distribution of various landforms: Open sloping land accounts for the largest proportion of the area and is distributed in Chuxiong and central Kunming. Deeply dissected canyons and U-shaped valleys are concentrated in western Yuxi and northwestern Chuxiong. The final output classification map should include heatmaps showing the boundaries, area proportions, and contribution of key factors for each landform type.

9. A complex mountainous terrain intelligent classification system based on dual-scale TPI optimization, characterized in that, The intelligent classification method for complex mountainous terrain based on dual-scale TPI optimization, as described in any one of claims 1 to 8, includes: The data preprocessing module is used to acquire ASTER GDEM data of the study area, divide the data into multiple terrain analysis units, extract the initial terrain factors of each unit, the terrain includes large-scale terrain location index, small-scale terrain location index, slope, and elevation, standardize the initial terrain factors to obtain the feature data of each factor, calculate the Pearson correlation coefficient and variance inflation factor between the feature data, determine the factor collinearity based on the Pearson correlation coefficient and variance inflation factor, and generate a collinearity diagnostic series. The factor screening module is used to screen and calibrate the factors in the collinearity diagnostic sequence. The screening and calibration includes high collinearity calibration and low collinearity calibration. The number of factors with high collinearity calibration is counted, the number of factors with low collinearity calibration is counted, and the optimal combination of terrain factors is determined based on the proportion of the two types of calibration. The model parameter tuning module is used to acquire historical geomorphic classification data of the study area, match the historical geomorphic classification data with the optimal factor combination, and determine whether to adjust the input parameters of the machine learning model based on the matching results. When it is determined that adjustment is needed, the primary influence coefficient of the model is determined based on the weight ratio of the terrain factor, and the target influence coefficient of the model is determined in combination with the number of sample points. The classification output module is used to determine the hyperparameter adjustment factors of random forest, extreme gradient boosting and deep neural network according to the model target influence coefficient, optimize the model parameters according to the adjustment factors, use the optimized model to classify the landform of the study area, and output the classification results.

Images