Micro-terrain recognition method and device fusing multi-source data and image processing technology

Abstract

The present application provides a kind of microrelief identification method and device fusing multi-source data and image processing technology, it is related to computer data processing technical field, including: the multi-source data of target area is collected;Multi-source data includes remote sensing image, vector data and terrain factor;Through microrelief identification model, the microrelief information of target area is determined by being input sample to the multi-source data of target area after identification;Microrelief identification model is obtained by following method training: establish microrelief training sample library, and the training sample is expanded by selecting terrain factor, sample enhancement processing and image processing mode;A plurality of initial models of different neural network structures are constructed, and the training sample of microrelief training sample library is used to train multiple initial models respectively, and after multiple rounds of dynamic optimization training strategy, model parameters are adjusted and optimized, and the performance index corresponding to each model is determined;Select the model that performance index meets preset condition as microrelief identification model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer data processing, and in particular to a micro-terrain recognition method and device fusing multi-source data and image processing technology. BACKGROUND

[0002] This section is intended to provide background or context to the embodiments of the application. The description herein does not constitute admission that the information provided herein is prior art.

[0003] With the continuous expansion of the power grid, the number of transmission lines passing through complex geographical environments and harsh weather conditions is increasing. Due to the diversity of topography and the wide distribution of rivers and lakes, many transmission lines need to pass through micro-terrain areas with complex landform morphologies such as high altitudes and gorges. These areas are significantly disturbed by topography to local meteorological factors, often causing transmission line icing phenomena, which seriously threaten the safe and stable operation of the power system. Icing disasters usually occur in extremely cold, high-altitude, and gorge areas with complex weather conditions. When the line passes through such micro-terrain areas, local temperature, humidity, wind speed, and other meteorological factors change dramatically under the action of topography, significantly increasing the icing risk and becoming an important inducement to power grid accidents. According to the formation mechanism of meteorological factors and the characteristics of the terrain, the micro-terrain icing area can be divided into five types: high mountain watershed type, pass type, terrain uplift type, gorge wind channel type, and water vapor increase type. Conducting classification and recognition research on micro-terrain areas can help to clarify the local icing causes and provide a basis for differentiated anti-icing measures, especially in ultra-high voltage transmission lines, which can improve the pertinence of disaster prevention design and the safety of power grid operation.

[0004] Current micro-terrain recognition mainly uses two methods: one is field investigation method, which collects landform feature information through field reconnaissance, and classifies micro-terrain types based on experience. The precision is high, but the labor cost is high, and the coverage is limited. The second is the terrain factor threshold method, which extracts slope, slope direction, and elevation difference based on DEM data, and classifies the area by setting empirical thresholds. This method has certain automation capability and is suitable for preliminary screening of large areas, but the subjective determination rule is strong and the robustness is poor, which is difficult to cope with complex and variable micro-scale landform environment, and is not suitable for high-precision micro-terrain automatic recognition requirements.

[0005] Therefore, it is urgent to propose a micro-terrain recognition method fusing multi-source geographic data and image processing technology, which introduces remote sensing image information and topographic factor image expression to improve the automation level and recognition accuracy of micro-terrain classification, and meet the actual needs of accurate recognition of complex micro-terrain in transmission line anti-icing work. SUMMARY

[0006] To solve the problems in the prior art, the present application provides a micro-terrain recognition method and device fusing multi-source data and image processing technology.The present application can solve the deficiencies of the prior art in recognition efficiency, accuracy, application range and robustness, and realizes effective recognition of complex micro-landform types by introducing image expression of remote sensing image information and terrain factors and combining a nonlinear discrimination model with multi-level feature extraction capability, thereby constructing an efficient and stable micro-terrain recognition process system suitable for complex landform regions.

[0007] In a first aspect of the embodiments of the present application, a micro-terrain recognition method fusing multi-source data and image processing technology is provided, which comprises:

[0008] Collecting multi-source data of a target region; wherein the multi-source data comprises remote sensing images, vector data and terrain factors;

[0009] Determining micro-terrain information of the target region by a micro-terrain recognition model through recognition of the multi-source data of the target region as input samples; wherein the micro-terrain recognition model is obtained by training through the following method:

[0010] Establishing a micro-terrain training sample library, expanding training samples by selecting terrain factors, sample enhancement processing and image processing methods; constructing multiple initial models with different neural network structures, training the multiple initial models respectively by training samples of the micro-terrain training sample library, adjusting and optimizing model parameters through multiple rounds of dynamic optimization training strategies, and determining performance indicators corresponding to each model; and selecting a model satisfying a preset condition as the performance indicators as the micro-terrain recognition model.

[0011] In a second aspect of the embodiments of the present application, a micro-terrain recognition device fusing multi-source data and image processing technology is provided, which comprises:

[0012] A data collection module for collecting multi-source data of a target region; wherein the multi-source data comprises remote sensing images, vector data and terrain factors;

[0013] A micro-terrain recognition module for determining micro-terrain information of the target region by a micro-terrain recognition model through recognition of the multi-source data of the target region as input samples;

[0014] The micro-terrain recognition module comprises:

[0015] A sample processing unit for establishing a micro-terrain training sample library, expanding training samples by selecting terrain factors, sample enhancement processing and image processing methods;

[0016] The model training unit is configured to construct initial models of different neural network structures, train the initial models by using training samples in the micro-terrain training sample library, adjust and optimize model parameters through dynamic optimization training strategies for multiple rounds, and determine performance indicators corresponding to each model; and select a model satisfying a preset condition in the performance indicators as the micro-terrain recognition model.

[0017] In a third aspect of the embodiments of the present application, a computer device is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the micro-terrain recognition method fusing multi-source data and image processing technology when executing the computer program.

[0018] In a fourth aspect of the embodiments of the present application, a computer readable storage medium is provided, which stores a computer program, and the computer program implements the micro-terrain recognition method fusing multi-source data and image processing technology when executed by a processor.

[0019] In a fifth aspect of the embodiments of the present application, a computer program product is provided, which includes a computer program, and the computer program implements the micro-terrain recognition method fusing multi-source data and image processing technology when executed by a processor.

[0020] The micro-terrain recognition method and device fusing multi-source data and image processing technology can realize efficient recognition of micro-terrain by fusing terrain factor image expression, terrain feature enhancement, and multi-scale feature fusion technology. The overall performance of the recognition system in terms of precision, efficiency, and robustness is effectively improved by automatically constructing a micro-terrain training sample library and implementing a multi-stage model optimization strategy. In particular, in the terrain data processing, the spatial interpolation and feature selection methods are introduced to enhance the expression ability of key terrain point and line information, significantly improving the extraction accuracy of micro-terrain features. At the same time, a multi-source data input system is constructed, and different terrain factor grayscale images are superimposed on remote sensing images to provide more rich geomorphic information input for the model. The model training stage adopts a step-by-step optimization mechanism and performance tuning based on precision, recall rate, and other evaluation indicators to ensure its stable generalization ability in complex terrain scenarios. The method shows higher accuracy in the recognition of micro-terrain regions in typical power transmission lines, effectively supports differentiated anti-icing and ice-resistant design, and improves the operation safety and reliability of power grids under extreme climate conditions, providing key support for ensuring the stable operation of energy transmission channels. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort.

[0022] Figure 1 is a micro-terrain recognition method flowchart of an embodiment of the present application, which fuses multi-source data and image processing technology.

[0023] Figure 2 is a micro-terrain recognition model training flowchart of an embodiment of the present application.

[0024] Figure 3 is a micro-terrain training sample database construction flowchart of an embodiment of the present application.

[0025] Figure 4 is a residual structure stacking method flowchart of an embodiment of the present application.

[0026] Figure 5 is a micro-terrain recognition model structure design and model training technology flowchart of an embodiment of the present application.

[0027] Figure 6 is a loss performance diagram of each model in the validation set of an embodiment of the present application.

[0028] Figure 7 is a micro-terrain recognition device architecture flowchart of an embodiment of the present application, which fuses multi-source data and image processing technology.

[0029] Figure 8 is a computer device structure flowchart of an embodiment of the present application. DETAILED DESCRIPTION

[0030] The principles and spirits of the present application will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are given only to enable those skilled in the art to better understand and implement the present application, and do not limit the scope of the present application in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0031] Those skilled in the art know that the embodiments of the present application can be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure can be embodied in the form of a complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0032] According to an embodiment of the present application, a micro-terrain recognition method and device integrating multi-source data and image processing technology are provided, relating to the technical field of computer data processing.

[0033] The principles and spirits of the present application will be explained in detail below with reference to several representative embodiments of the present application.

[0034] Figure 1 Figure 1 is a flowchart of a micro-terrain recognition method integrating multi-source data and image processing technology according to an embodiment of the present application. Figure 1 As shown in the figure, the method comprises the following steps.

[0035] S101, collecting multi-source data of a target area; wherein the multi-source data comprises remote sensing images, vector data and terrain factors;

[0036] S102, determining micro-terrain information of the target area by a micro-terrain recognition model through recognition of the multi-source data of the target area as input samples;

[0037] The micro-terrain recognition model is obtained by the following method:

[0038] A micro-terrain training sample library is established, and the training samples are expanded by selecting terrain factors, sample enhancement processing and image processing methods; a plurality of initial models of different neural network structures are constructed, and the plurality of initial models are trained by the training samples of the micro-terrain training sample library, respectively; the model parameters are adjusted and optimized through multiple rounds of dynamic optimization training strategies, and the performance indicators corresponding to each model are determined; and the model satisfying the preset condition in the performance indicators is selected as the micro-terrain recognition model.

[0039] The present application integrates terrain factor image expression, terrain feature enhancement and multi-scale feature fusion technology, and realizes efficient recognition of micro-terrain. By automatically constructing a micro-terrain training sample library and implementing a multi-stage model optimization strategy, the overall performance of the recognition system in terms of accuracy, efficiency and robustness is effectively improved. In particular, in the terrain data processing, the spatial interpolation and feature selection method is introduced to enhance the expression ability of the key terrain point line information, and the extraction accuracy of the micro-terrain feature is significantly improved. At the same time, a multi-source data input system is constructed, and different terrain factor grayscale images are superimposed on remote sensing images to provide more rich geomorphic information input for the model. In the model training stage, a step-by-step optimization mechanism is adopted, and the performance is optimized based on the precision, recall rate and other evaluation indicators to ensure its stable generalization ability in complex terrain scenes. The method shows higher accuracy in the recognition of micro-terrain areas of typical power transmission lines, can effectively support differentiated anti-icing and ice-resistant design, and improve the operation safety and reliability of the power grid under extreme climate conditions, providing key support for ensuring the stable operation of energy transmission channels.

[0040] In order to make the micro-terrain recognition method combining the fusion of multi-source data and image processing technology more clearly explained, the following will be described in detail in conjunction with specific embodiments.

[0041] In an embodiment, when establishing the micro-terrain training sample library, the training samples can be expanded in the following one or more ways when selecting terrain factors, sample enhancement processing and image processing methods:

[0042] Selecting terrain factors, quantitatively analyzing the terrain factors, and retaining terrain factors meeting the preset requirements;

[0043] Performing sample enhancement processing based on digital elevation model, digital orthophoto map and SHP format data;

[0044] Preprocessing the training samples, and the preprocessing methods at least include projection, cropping and generating a grayscale image.

[0045] Specifically, when constructing the micro-terrain training sample library, first, the selection and extraction of terrain factors, quantitative analysis of terrain factors and finding data sources are started, then the DEM (digital elevation model), DOM (digital orthophoto map) and SHP auxiliary data are used for sample enhancement and data preprocessing (including projection, cropping and grayscale image generation), the samples are expanded through these steps, and finally the sample database is constructed for further research and analysis of micro-terrain.

[0046] SHP is a vector data format used to store geographic spatial data and attribute information, and is widely used in the field of GIS (Geographic Information System). The SHP file consists of three core files:.shp: stores the location and shape of graphic data (points, lines, polygons, etc.);.shx: index file of graphic data, accelerates data retrieval;.dbf: stores attribute information of geographic features (such as name, type, etc.).

[0047] For the sample expansion method, the specific method of selecting terrain factors, quantitatively analyzing the terrain factors, and retaining terrain factors meeting the preset requirements includes:

[0048] According to the variance inflation factor corresponding to each terrain factor, the variance inflation factor is calculated by the following formula:

[0049]

[0050] In the formula, is the square of the complex correlation coefficient of the jth variable to other variables; VIF jThe VIF value corresponding to the jth variable is represented, and the larger the VIF value, the higher the degree of interpretation of the variable by other variables, and the stronger the redundancy; the variables with VIF values greater than a preset threshold are removed, and the variables with VIF values less than or equal to the preset threshold are retained.

[0051] For the sample expansion method, the specific method for sample enhancement processing based on the digital elevation model, the digital orthographic image and the SHP format data includes:

[0052] The spatial coordinates of the grid center point of the digital elevation model are taken as the characteristic variable, and the elevation value is taken as the dependent variable to construct a random forest regression model; a plurality of random points are inserted in the original grid, and the spatial coordinates are taken as the input to predict the elevation value through spatial interpolation to enhance the digital elevation model; the elevation mean difference between the grid center point and the random point inside the grid is calculated as an importance measurement index of the grid unit, and the grid center point with an importance exceeding a preset importance threshold is screened out and extracted as a candidate terrain feature point;

[0053] According to the candidate terrain feature point, a spatial distribution map of feature points and lines is drawn, and the digital elevation model under the original resolution is reconstructed; all grid points in the digital elevation model under the original resolution are taken as checkpoints to evaluate the error statistical index of the reconstructed digital elevation model, and the matching condition of the feature line is analyzed;

[0054] According to the spatial distribution of the feature points, the reconstructed error statistical index, and the matching condition of the feature line, the sample enhancement processing is evaluated to determine the sample enhancement effect.

[0055] Further, the sample enhancement processing based on the digital elevation model, the digital orthographic image and the SHP format data includes:

[0056] The CART decision tree is used to divide the training sample space, and the minimum elevation deviation sum of squares of two sub-regions is taken as the division standard, and the minimum elevation deviation sum of squares is calculated by the following formula:

[0057]

[0058] In the formula, SS represents the minimum elevation deviation sum of squares of two sub-regions; z i is the elevation value of the sampling point in the region; is the average value of the elevation values of all sampling points in the region;

[0059] Based on the divided training sample space, random sampling is performed from the original terrain sampling point data to generate multiple training sample sets; wherein the capacity of each training sample set remains the same as the original data set, the plane coordinates of the grid center point are taken as the input features, and the elevation is taken as the prediction target, and the CART decision tree model is trained; through recursive division of the plane coordinate space, the sample data is divided into leaf nodes to form a complete regression tree; the regression tree is used as a simulation model of the terrain surface;

[0060] By repeating the above sampling and tree building process, multiple regression trees are constructed to form a random forest model composed of multiple regression trees; wherein for random points in the same grid cell, the random forest provides multiple predicted elevation values, and the mean of the multiple predicted elevation values is taken as the final elevation estimate of the corresponding random point.

[0061] In an embodiment, a micro-terrain training sample library is established, and the training samples are expanded through selection of terrain factors, sample enhancement processing and image processing methods, including:

[0062] Based on stratified random sampling, a first number of training areas and a first number of test areas are selected;

[0063] For the training area and the test area, by fusing the pre-set resolution digital orthophoto map, semi-automatic boundary extraction is used to assist manual interpretation to demarcate the sample range to obtain samples and labels; wherein the samples of the training area and the test area are projected to a target coordinate system, and bicubic interpolation and resampling processing are performed to construct a digital orthophoto map containing RGB three bands and a feature data set of multiple terrain factors;

[0064] For the digital orthophoto map and the feature data set of the terrain factor, a sliding window and SHP data are used to determine the position and range of the water body to generate training samples; wherein after the training sample construction is completed, a part of the training image is subjected to data enhancement processing such as rotation or inversion.

[0065] In an embodiment, for the micro-terrain recognition model, a basic backbone network with ResNet18 as the initial model can be constructed, and ACON is used as the activation function in the residual block of the initial model,

[0066] The feature pyramid network is taken as the neck of the initial model, and the convolution layer is taken as the head of the initial model.

[0067] In the structure design and model training of the micro-terrain recognition model, an advanced micro-terrain recognition model structure construction technology is adopted, a model training method based on a multi-stage iterative optimization training strategy is adopted, and a micro-terrain recognition exclusive intelligent reasoning model is constructed. The model takes ResNet18 as the basic backbone network for feature extraction, introduces ACON activation function to enhance the feature expression ability, uses feature pyramid as the neck for multi-scale feature fusion, and uses convolution layer as the head to output the final recognition result. Through the dynamic optimization training strategy of multiple rounds, the model parameters are continuously adjusted and optimized to improve the recognition accuracy and generalization ability of the model to micro-terrain features, and finally realize efficient and accurate micro-terrain recognition.

[0068] In an embodiment, multiple initial models of different neural network structures are constructed, and the multiple initial models are trained by training samples of the micro-terrain training sample library. Through multiple rounds of dynamic optimization training strategies, the model parameters are adjusted and optimized, and the performance indicators corresponding to each model are determined. The model whose performance indicator meets the preset condition is selected as the micro-terrain recognition model, and the specific method includes:

[0069] An initial model corresponding to each of the traditional convolutional neural network AlexNet, VGGNet, the classic deep network ResNet, and the introduction of attention mechanism ACON, random forest feature enhancement RF, RF+ResNet, ACON-ResNet, and RF+ACON-ResNet is constructed.

[0070] After training and testing the multiple initial models, quantitative analysis and qualitative evaluation are performed on each model to obtain corresponding performance indicators. The model with the highest performance indicator ranking is selected as the micro-terrain recognition model.

[0071] The performance indicators at least include precision, recall, F1 score, and prediction accuracy.

[0072] Compared with the prior art, the present application has at least the following improvements:

[0073] 1. Terrain feature enhancement and sample construction method:

[0074] The present application proposes a spatial feature enhancement method combined with high-precision terrain factors. By mining and strengthening the key feature points and feature lines in the digital elevation model, the expression ability of the terrain sample in terms of local fluctuation change and slope difference is improved. In the sample construction process, a spatial interpolation and feature selection strategy based on machine learning is adopted, which effectively improves the representativeness and difference of the training data, and provides more discriminative input basis for the recognition task of complex micro-terrain.

[0075] 2. Multi-source data fusion:

[0076] The data processing mode of fusing remote sensing images, vector data and various terrain factors (such as elevation, slope and cutting depth) is constructed, and information of different sources and different scales is unified modeling in the same coordinate system. Through the image conversion and linkage expression of multi-source data, the fine-grained perception of complex micro-terrain form is realized, the recognition ability of the model to micro-landform change is improved, and the adaptability of the overall system in complex geographical environment is enhanced.

[0077] 3. Intelligent discrimination method for micro-terrain:

[0078] A network structure with nonlinear expression ability and deep semantic abstraction ability is constructed, and the accuracy and stability in the micro-terrain recognition task are effectively improved by introducing a cross-layer connection mechanism and a multi-scale feature fusion strategy. The structure can adaptively extract micro-terrain features at different scales, enhance the ability to capture complex boundaries and detail changes, and is suitable for high-precision micro-terrain classification requirements in large-scale and sample-scarce scenarios.

[0079] It should be noted that although the operations of the method of the present application are described in a specific order in the above embodiments and drawings, this does not require or imply that the operations must be performed in this specific order, or that all of the shown operations must be performed to achieve the desired result. Additionally or alternatively, certain steps can be omitted, a plurality of steps can be combined into one step, and / or one step can be divided into a plurality of steps.

[0080] The micro-terrain recognition method combining multi-source data and image processing technology will be further described in detail below in conjunction with a specific embodiment.

[0081] Reference Figure 2 The overall schematic diagram of the training process of the micro-terrain recognition model of a specific embodiment of the present application is shown in FIG. 1. Figure 2 As shown in FIG. 1, after constructing a micro-terrain training sample database, designing a micro-terrain recognition model structure, training a model, and comparing and verifying the model, a micro-terrain recognition model is constructed, and micro-terrain recognition is realized based on the model. The specific process includes:

[0082] S201, constructing a micro-terrain training sample database. The main steps include terrain factor selection, terrain feature enhancement and extraction, and sample construction.

[0083] S202, micro-terrain recognition model structure design and model training. The main steps include constructing a micro-terrain recognition model and training a micro-terrain recognition model.

[0084] S203, model comparison and verification. The main steps include performance index qualitative evaluation and model loss comparison.

[0085] Determining the micro-terrain recognition result based on the micro-terrain recognition model.

[0086] Detailed explanation is made for each step:

[0087] S201, construct a micro-terrain training sample database.

[0088] From the selection and extraction of terrain factors, quantitative analysis of terrain factors, and finding data sources, then using DEM (Digital Elevation Model), DOM (Digital Orthophoto Map) and SHP auxiliary data for sample enhancement and data preprocessing (including projection, cutting and gray map generation), through these steps to expand the sample, finally form a sample database for further research and analysis of micro-terrain, the specific process relationship is shown in Figure 3 .

[0089] Sub-step one, select, extract terrain factors and perform quantitative analysis, find data sources.

[0090] This step mainly includes preliminary screening, extraction and quantitative analysis of terrain factors, and correlation test and determination of final factors.

[0091] First, according to the existing theoretical knowledge, the preliminary selected terrain factors include two categories of macro-terrain factors and micro-terrain factors. Among them, the macro-factor is used to reflect the overall terrain characteristics of the region, including terrain relief degree, elevation variation coefficient, and surface roughness; the micro-factor describes the local surface details including average slope, slope variability, profile curvature and plane curvature, etc. Terrain information index, the above factors are extracted through DEM, and Z-score is used for standardization processing to eliminate the influence of dimension difference on analysis.

[0092] Second, the micro-terrain factors are extracted and analyzed. The first law of geography points out that spatial adjacent geographical elements have correlation. The multiple collinearity may be caused by the physical correlation between terrain factors (such as slope variability depending on slope), which affects the stability of subsequent modeling. Therefore, there may be a certain degree of mutual correlation, mutual restriction and mutual influence between different terrain factors. Therefore, it is necessary to perform quantitative correlation coefficient analysis on these terrain factors which describe the surface relief and variation characteristics from different angles, in order to eliminate redundant and high information overlap factors, which not only guarantees the accuracy of landform type identification, but also reduces the data dimension and calculation burden. The invention uses variance inflation factor (VIF) to quantify the degree of collinearity, and the specific calculation formula is as follows:

[0093]

[0094] In the formula, is the square of the multiple correlation coefficient of the jth variable to other variables; VIF j VIF value of the jth variable. The larger the VIF value, the higher the degree of the variable explained by other variables, and the stronger the redundancy.

[0095] The calculation process is as follows: a regression model of all terrain factors is constructed, each variable is sequentially taken as the dependent variable, and the remaining variables are taken as the independent variables, and R 2 The VIF value of each variable is calculated. According to the definition of VIF, when VIF is 5, the corresponding variable can be explained 80% of the variance by other variables (i.e. only 20% independent information is retained), and the variance inflation is controlled while retaining most of the independent information. Therefore, the threshold is set to 5, and three factors of surface roughness, average slope variability and coefficient of variation of elevation are removed, and four terrain factors of average elevation, average slope, surface cutting depth and terrain relief amplitude are obtained.

[0096] The VIF calculation results are shown in Table 1.

[0097] Table 1 VIF value of terrain factor

[0098]

[0099] Through the screening of VIF, it is determined that the above four terrain factors can eliminate the collinearity interference and improve the stability of the model, and are used for subsequent automatic recognition and deep learning modeling of micro-terrain types. Finally, DOM and DEM data are selected as the model input data source.

[0100] Substep two, terrain feature enhancement and extraction.

[0101] Because the grid data of DEM is huge and contains some redundant information, in actual analysis and application, it is usually necessary to extract the main terrain features with rich information and strong representation from DEM, so as to accurately express the real terrain surface with a small amount of high-quality feature data, thereby improving the data processing efficiency and storage performance. Therefore, how to accurately and efficiently extract key terrain features has important significance for reducing terrain data redundancy, meeting the demand of multi-scale spatial data, and constructing an efficient spatial database. Especially in micro-terrain recognition, high extraction accuracy requirements are put forward for feature lines (such as valley lines, ridge lines, etc.) and feature points (such as mountain top points, saddle points, mountain points, etc.). Therefore, the present application proposes to enhance the extraction accuracy of feature points and feature lines by enhancing the samples of DEM data.

[0102] The specific method is as follows: taking the spatial coordinates (X, Y) of the center points of the DEM grid as characteristic variables and taking the elevation value (Z) as the dependent variable, a random forest regression model is constructed. By inserting a plurality of random points in the original grid and taking the spatial coordinates of the random points as input, the trained model is used for spatial interpolation to predict the elevation value, so that the DEM data is enhanced. On this basis, the mean difference of the elevations of the grid center points and the random points in the grid is calculated as an importance measurement index of the grid unit, and the grid units with high importance are screened out and the center points thereof are extracted as candidate terrain feature points.

[0103] Subsequently, a terrain feature extraction experiment based on the above method is carried out, a spatial distribution map of feature points and lines is drawn, and the original resolution DEM is reconstructed based on the features. All grid points in the original DEM are taken as check points to evaluate the error statistical indicators of the reconstructed DEM.

[0104] The method of the application is compared with the maximum Z tolerance method and the important point method, and is comprehensively evaluated from three dimensions of feature point spatial distribution, reconstruction error statistical indicators and feature line matching, and the effectiveness of the method is tested by combining qualitative and quantitative methods.

[0105] In the model construction process, CART (Classification and Regression Tree) decision tree is used to divide the training sample space, and the minimum elevation deviation square sum (SS) of two sub-regions is used as the division standard, and the calculation formula is as follows:

[0106]

[0107] In the formula, SS represents the minimum elevation deviation square sum of two sub-regions; z i is the elevation value of the sampling point in the region; is the average value of the elevation values of all sampling points in the region;

[0108] The specific operation is as follows: based on the divided training sample sub-regions, a plurality of training sample sets are generated by random sampling with replacement from the original terrain sampling point data. The capacity of each sample set is the same as that of the original data set, the spatial coordinates (X, Y) of the DEM grid center points are taken as input features, and the elevation is taken as a prediction target (dependent variable), and a CART decision tree model is trained. Through recursive division of the plane coordinate space, the elevation impurity in each sub-region is minimized, and finally the sample data is divided into various leaf nodes to form a complete regression tree. The regression tree can be used as a simulation model of the terrain surface.

[0109] To improve the prediction accuracy, the application further constructs a random forest model. By repeating the above sampling and tree building process, several regression trees are constructed to form a random forest model composed of multiple regression trees. For random points in the same grid cell, the random forest can provide multiple predicted elevation values, and the mean value is taken as the final elevation estimate of the point. In this way, the sample enhancement of DEM data and the accurate extraction of terrain feature points are realized.

[0110] Sub-step three, sample construction.

[0111] Firstly, 60 training areas and 12 test areas are selected based on stratified random sampling to ensure the spatial representativeness and quantitative balance of the five types of micro-terrain samples. By fusing 0.5-meter resolution DOM images and using semi-automatic boundary extraction assisted by manual interpretation, the sample range is accurately delineated to obtain the sample and label. All data is projected into the UTM-WGS84 coordinate system and resampled by bicubic interpolation to ensure that the spatial resolution is aligned, and a feature dataset containing DOM images (RGB three bands) and the remaining four terrain factors (average elevation, average slope, surface cutting depth, and terrain relief amplitude) is constructed. A sliding window of 512*512 size is used, and SHP data such as water bodies are superimposed to clearly define the location and range of water bodies, and training samples are generated. After the sample construction is completed, some training images are rotated, inverted, and other data augmentation processing is performed to increase the diversity of the samples.

[0112] S202, micro-terrain recognition model structure design and model training.

[0113] Sub-step one, construct a micro-terrain recognition model.

[0114] A ResNet18 model is constructed as the basic backbone network. ResNet18 is a deep residual network that can effectively extract image features and alleviate the gradient vanishing problem in deep network training. In the context of a large study area and limited available sample size, selecting this structure as the basic model can balance the recognition accuracy, model training efficiency, and generalization ability.

[0115] Secondly, ACON is used as the activation function in the residual block. ACON activation function is mainly located after the second convolution layer of the ResNet model residual structure. Replace the ACON activation function with the ReLU activation function, so that the residual calculation can adaptively learn the linear or nonlinear ACON activation function according to the input feature map. At this time, the residual structure stacking method is as shown in Figure 4 .

[0116] In Figure 4 , the input (Input) is the feature map to be processed (such as image features, terrain factor features).

[0117] First 3x3 convolution (3x3conv): performs spatial feature extraction on the input, keeping the number of channels / dimensions matching the input (for ease of subsequent residual connections).

[0118] ACON activation: an adaptive activation function that replaces traditional ReLU, dynamically adjusting the activation mode (linear / non-linear) based on the input, enhancing the ability to capture complex features such as micro-terrain textures.

[0119] Second 3x3 convolution (3x3conv): further extracts features, outputting feature maps consistent with the input dimensions.

[0120] Residual connection (⊕): adds the original input (Input) directly to the output of the second convolution (element-wise addition), implementing residual learning (network learning, the difference between input and output, rather than absolute output), alleviating deep network gradient vanishing.

[0121] ReLU activation (relu): applies a non-linear transformation to the result of the residual connection, enhancing the model's expression ability.

[0122] Again, the feature pyramid network is used as the neck of the model. The feature pyramid network can extract features from different levels and construct a pyramid structure to fuse multi-scale information, which helps to capture targets of different sizes.

[0123] Finally, the convolutional layer is used as the head of the model. The convolutional layer is used to generate the final output, where the output channel number of the convolutional layer is usually matched with the number of target categories.

[0124] Sub-step two, micro-terrain recognition model training.

[0125] According to the previously constructed micro-terrain type training sample database.

[0126] Before starting model training, the network structure and configuration of the model need to be built, and some model training parameters are preset, including activation function (ACON), loss function (cross-entropy loss), optimization function (Adam), learning rate (lr), batch size (batch), training rounds (epoch), etc. Secondly, the model is trained using the preset parameters. This step involves learning and optimization of the model on the sample library.

[0127] Again, during the training process, the indicators on the validation set and the weights and training state of the model are recorded for evaluating the performance of the model. Whether the indicators on the validation set reach the optimal state indicators are respectively: precision, recall, and the comprehensive score of balanced precision and recall (F1-score, F1 score). The overall accuracy (OA, Overall Accuracy) refers to the correct classification samples divided by the total samples, which measures the final overall accuracy, and the calculation formula is as follows:

[0128]

[0129]

[0130]

[0131]

[0132] In the formula, TP represents the true class (true example), which is the number of samples of a certain type of micro-terrain correctly identified by the model as the type in the present application scenario, and its meaning is to identify the correct result of the terrain type;

[0133] FP represents the false positive class (false positive example), which is the number of samples of a certain type of micro-terrain incorrectly identified by the model as the type in the present application scenario;

[0134] FN represents the false negative class (false negative example), which is the number of samples of a certain type of micro-terrain incorrectly missed (not identified as the type) by the model in the present application scenario.

[0135] TN represents the true negative class (true negative example), which is the number of samples of a certain type of micro-terrain correctly identified by the model as "not the type" in the present application scenario.

[0136] Precision represents the precision, which is a measure of the accuracy of the model. Recall represents the recall, which is a measure of the ability of the model to extract information.

[0137] F1 score represents the comprehensive score of balanced precision and recall, and in the case of equal importance of precision and recall, the harmonic mean of the two is calculated. F1 score is a comprehensive performance benchmark, especially suitable for micro-terrain identification scenarios with balanced samples. When it is necessary to control "wrong judgment" and "omission" at the same time, F1 score is preferred.

[0138] OA represents the overall accuracy, and the proportion of correct prediction results to all prediction samples. OA is suitable for scenarios with balanced sample distribution (e.g., training samples can cover multiple types of micro-terrain and are balanced in quantity), and can quickly reflect the overall effect of the model. If a certain type of sample accounts for a very high proportion (e.g., non-micro-terrain samples account for 90%), even if the model is only accurate for this type of classification, OA will be high, at which time F1 and other indicators need to be combined for further judgment.

[0139] According to the above indicators, if the indicators on the verification set reach the optimal value, the weight of the current model is saved as the final model weight. If the indicators on the verification set do not reach the optimal value, the model parameters need to be adjusted, and the model training step is returned to for the next round of training.

[0140] Reference Figure 5 , a micro-terrain recognition model structure design and model training technical flowchart of an embodiment of the present application. As shown in Figure 5 , the overall process of model initialization, training optimization, and output terrain classification is exemplarily shown, and each link is deeply corresponding to the technical scheme of the micro-terrain recognition model structure design and model training in the document, as follows:

[0141] The first stage is data training, and the model parameters are adjusted through iterative training to achieve optimal recognition performance.

[0142] Select the backbone network ResNet18:

[0143] The basic feature extraction architecture of the model is determined as ResNet18 (deep residual network). ResNet18 can effectively alleviate the gradient disappearance problem in deep network training, and in the micro-terrain recognition scene with large area and limited sample quantity, the recognition accuracy, training efficiency, and generalization ability can be considered.

[0144] Modify the ACON activation function of the residual block:

[0145] Replace the traditional ReLU activation function in the ResNet18 residual block with the ACON activation function (after the second convolutional layer in the residual structure). ACON can allow residual calculation to adaptively learn linear or nonlinear features, enhance the model's feature expression ability for subtle textures of micro-terrain (such as slope changes and surface cutting depth differences), and reduce redundant information interference.

[0146] Determine the loss function and optimization function:

[0147] Configure the core calculation rules for model training, the loss function uses cross-entropy loss (used to measure the difference between the model prediction result and the true label), and the optimization function uses Adam (used to efficiently update model parameters and accelerate convergence).

[0148] Parameter initialization:

[0149] For the network modules of ResNet18, such as convolutional layers, fully connected layers, etc., set the initial parameter values (such as weights, biases) to lay the foundation for subsequent training.

[0150] Iterative training (mixed precision):

[0151] Based on the constructed micro-terrain training sample library, use mixed precision training to update parameters multiple times. Mixed precision training can improve training speed (reduce memory usage) while ensuring calculation accuracy, matching the multi-stage iterative optimization training strategy in the document.

[0152] Provide test set verification to determine whether it is optimal:

[0153] Use an independent test set (12 test area samples) to evaluate the current model performance and determine whether it has reached the optimal state. The evaluation criteria include precision, recall, F1 score, and overall accuracy (OA) as core indicators. If the indicators have not reached the optimal state, return to the iterative training step to continue adjusting the parameters. If the indicators have reached the optimal state, proceed to the next step.

[0154] Preserve model weights:

[0155] Save the parameters of the optimal model (weight file) as the core basis for the subsequent prediction phase, completing the transition from the training phase to the prediction phase.

[0156] The second phase is data prediction (model application), which aims to classify and identify new unknown terrain data.

[0157] Input the data to be predicted:

[0158] Data source: Multi-source geographic data (such as 0.5-meter resolution DOM images, 4 core terrain factors, average elevation, average slope, surface cutting depth, and terrain relief amplitude) that are homologous to the training samples and have been processed for projection unification and resolution alignment.

[0159] Model inference:

[0160] Load the preserved optimal model weights, input the data to be predicted into the model, and complete the automatic identification of terrain features through the feature extraction of ResNet18, the adaptive activation of ACON, and the multi-scale fusion of the feature pyramid.

[0161] Output terrain classification:

[0162] Output the micro-terrain types of the area to be predicted (corresponding to the 5 types of icing micro-terrain mentioned in document 1.1: high mountain watershed type, pass type, terrain uplift type, canyon wind channel type, and water vapor increase type), achieving the ultimate goal of high-precision micro-terrain identification.

[0163] S203, model comparison verification.

[0164] To explore the performance of different neural network structures in micro-terrain recognition tasks, the present application selects six typical models for comparison experiments, including traditional convolutional neural networks AlexNet and VGGNet, classic deep network ResNet, and improved models RF+ResNet, ACON-ResNet and RF+ACON-ResNet with attention mechanism (ACON) and random forest feature enhancement (RF).

[0165] To verify the performance of various models in micro-terrain recognition tasks, the present application selects six typical convolutional neural network architectures for comparison experiments, including traditional networks AlexNet and VGGNet, deep residual network ResNet, and improved models RF+ResNet, ACON-ResNet and RF+ACON-ResNet based on ResNet. All models are trained and evaluated on the same training set and test set, and their performance is comprehensively quantified and analyzed for qualitative evaluation using multiple indicators such as precision (Precision), recall (Recall), F1 score, overall accuracy (OA), etc. The specific performance is shown in Table 2.

[0166] Table 2 Specific performance of each model

[0167]

[0168] Referring to Table 2, the traditional models (AlexNet and VGGNet) have obvious limitations in shallow feature extraction, especially when facing terrain categories with fuzzy boundaries or subtle texture changes, the recognition accuracy is low, and the F1 score is only 0.514 and 0.518, which forms a bottleneck for overall classification performance. The ResNet model with deep structure enhances the feature extraction capability, significantly improves the classification effect, and the F1 score reaches 0.62, and the overall accuracy is 63.5%.

[0169] RF+ResNet with RF module effectively enhances the perception of terrain fine-grained structure through multi-scale feature reconstruction mechanism, and under the premise of keeping the change of model parameter quantity small, the F1 score is improved to 0.713, and the recall rate is also enhanced. On the other hand, ACON-ResNet adopts an adaptive activation mechanism to improve the nonlinear feature expression capability and reduce redundant feature interference, with Precision and F1 score reaching 70.3% and 0.693 respectively, and the overall performance is more balanced.

[0170] The RF+ACON-ResNet model, which combines the two improvement strategies, performs best across all metrics: precision reaches 82.1%, recall is 80.2%, F1 score is as high as 0.811, and overall accuracy (OA) is improved to 81.4%. In terms of inference efficiency, the model's inference time per image is only 14ms, achieving a good trade-off between speed and accuracy compared to traditional models.

[0171] Figure 6 This is a schematic diagram illustrating the loss performance of various models on the validation set according to a specific embodiment of the present invention. Figure 6 In this context, an epoch represents the training process of going through all samples in the training dataset once (and only once). Within an epoch, the training algorithm inputs all samples into the model in a predefined order for forward propagation, loss calculation, backpropagation, and parameter updates.

[0172] refer to Figure 6 All models showed a rapid decline in loss before epoch 5, with the difference between models maximizing between epochs 10 and 15. VGGNet performed the worst, exhibiting significant overfitting after epoch 10; AlexNet showed a unique "step-like decline," with a sharp drop every 5 epochs. Both models showed a linear decline in loss curves, lacking a rapid convergence phase. Due to its large parameter count (138M), VGGNet experienced the slowest initial loss decline, while AlexNet essentially stagnated after epoch 10. ResNet performed slightly better than the other two models, breaking the critical threshold 5 epochs earlier. ACON-ResNet outperformed RF+ResNet in the early stages (epochs 1-5) (loss 0.64 vs 0.66), but its convergence speed slowed down in later stages, allowing it to be overtaken. The ACON-ResNet curve exhibited a typical exponential decline, with a lower initial slope than RF+ResNet. It experienced a brief plateau at epoch 10 (loss ≈ 0.34), where the RF+ACON-ResNet model performed best, with its loss dropping below 0.3 at epoch 10 (compared to RF+ResNet's 0.73), demonstrating the fastest feature extraction capability. Furthermore, it only required 12 epochs to reach a stable loss of 0.08, 40% faster than other models. It was also the only model to break the 0.3 loss threshold before epoch 10. The RF+ACON-ResNet model exhibited the lowest validation loss at all stages of training, demonstrating optimal generalization ability. Compared to other classic network architectures, this model effectively improves the ability to identify micro-topographic features by combining random forest data augmentation with ACON activation mechanism, providing stronger expressiveness and stability for learning complex terrain.

[0173] In summary, the RF+ACON-ResNet realizes structural enhancement in two key aspects of input feature reconstruction and activation function optimization, significantly improves the classification performance of micro-terrain recognition while maintaining a low inference cost, is especially suitable for high-difficulty scenes with complex landform structures and fuzzy boundaries, and has strong engineering promotion value and practical application potential.

[0174] The present application proposes to fuse multi-source geographic data and image processing technology for micro-terrain recognition, introduces image expression of remote sensing image information and terrain factors, combines a nonlinear discriminant model with multi-level feature extraction capability, and realizes effective recognition of complex micro-landform types.

[0175] After introducing the method of the exemplary embodiments of the present application, next, with reference to Figure 7 The micro-terrain recognition device of the exemplary embodiments of the present application fusing multi-source data and image processing technology is introduced.

[0176] The implementation of the micro-terrain recognition device fusing multi-source data and image processing technology can refer to the implementation of the above method, and the repeated parts will not be described again. The term "module" or "unit" used below can be a combination of software and / or hardware that realizes the predetermined function. Although the device described in the following embodiments is preferably realized in software, the realization of hardware, or a combination of software and hardware is also possible and is conceived.

[0177] Based on the same inventive concept, the present application also proposes a micro-terrain recognition device fusing multi-source data and image processing technology, as shown in Figure 7 The device comprises:

[0178] The data acquisition module 710 is configured to acquire multi-source data of a target area; wherein the multi-source data comprises remote sensing images, vector data and terrain factors;

[0179] The micro-terrain recognition module 720 is configured to take the multi-source data of the target area as input samples, and determine the micro-terrain information of the target area through a micro-terrain recognition model;

[0180] The micro-terrain recognition module comprises:

[0181] The sample processing unit 721 is configured to establish a micro-terrain training sample library, and expand the training samples by selecting terrain factors, sample enhancement processing and image processing methods;

[0182] The model training unit 722 is configured to construct initial models of a plurality of different neural network structures, train the plurality of initial models respectively by using the training samples in the micro-terrain training sample library, adjust and optimize model parameters through a plurality of rounds of dynamic optimization training strategies, and determine a performance index corresponding to each model; and select a model satisfying a preset condition in the performance index as the micro-terrain recognition model.

[0183] It should be noted that, although several modules of the micro-terrain recognition device integrating multi-source data and image processing technology are mentioned in the foregoing detailed description, such division is merely exemplary and not mandatory. In fact, according to the embodiments of the present application, the features and functions of two or more modules described above can be embodied in one module. Conversely, the features and functions of one module described above can be further divided into a plurality of modules.

[0184] Based on the foregoing inventive concept, as shown in Figure 8 The present application also proposes a computer device 800, which comprises a memory 810, a processor 820, and a computer program 830 stored in the memory 810 and executable on the processor 820, wherein the processor 820 implements the foregoing micro-terrain recognition method integrating multi-source data and image processing technology when executing the computer program 830.

[0185] Based on the foregoing inventive concept, the present application proposes a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement the foregoing micro-terrain recognition method integrating multi-source data and image processing technology.

[0186] Based on the foregoing inventive concept, the present application proposes a computer program product, which comprises a computer program, wherein the computer program is executed by a processor to implement the foregoing micro-terrain recognition method integrating multi-source data and image processing technology.

[0187] The micro-terrain recognition method and device fusing multi-source data and image processing technology can realize efficient recognition of micro-terrain by fusing topographic factor image expression, topographic feature enhancement and multi-scale feature fusion technology. Through automatic construction of a micro-terrain training sample library and implementation of a multi-stage model optimization strategy, the overall performance of the recognition system in terms of precision, efficiency and robustness is effectively improved. In particular, in the topographic data processing, the spatial interpolation and feature screening method is introduced to enhance the expression ability of key topographic point and line information, and significantly improve the extraction accuracy of micro-terrain features. At the same time, a multi-source data input system is constructed, and different topographic factor grayscale maps are superimposed on remote sensing images to provide more abundant geomorphic information input for the model. In the model training stage, a step-by-step optimization mechanism is adopted, and performance tuning is performed based on accuracy, recall rate and other evaluation indicators to ensure stable generalization ability of the model in complex terrain scenes. The method has higher accuracy in recognizing micro-terrain areas of typical power transmission lines, can effectively support differentiated anti-icing and ice-resistant design, and improve the operation safety and reliability of the power grid under extreme climate conditions, thereby providing key support for ensuring the stable operation of energy transmission channels.

[0188] The acquisition, storage, use, processing and the like of data in the technical solutions of the present application comply with relevant provisions of laws and regulations.

[0189] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, apparatus, or computer program products. Accordingly, the application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) having computer usable program code embodied in the medium.

[0190] The present application is described with reference to flowcharts and / or block diagrams of the method and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts 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, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 An apparatus for performing each flow or multiple flows and / or blocks Figure 1 An apparatus for performing each flow or multiple flows and / or blocks

[0191] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the flow Figure 1 of the flow or flows and / or blocks Figure 1 of the block or blocks specified in the flow.

[0192] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 of the flow or flows and / or blocks Figure 1 of the block or blocks specified in the flow.

[0193] Finally, it should be noted that the above-described embodiments are merely exemplary implementations of the present application, and are used to explain the technical solutions of the present application, but are not intended to limit the present application. The protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can make modifications or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some of the technical features, without departing from the technical scope disclosed by the present application. These modifications, changes or replacements do not cause the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A micro-terrain recognition method fusing multi-source data and image processing technology, characterized in that, The method comprises: Collecting multi-source data of a target area; wherein the multi-source data comprises remote sensing images, vector data and terrain factors; Through a micro-terrain identification model, the multi-source data of the target area is taken as an input sample to determine the micro-terrain information of the target area; wherein the micro-terrain identification model is obtained by training the following method: A micro-terrain training sample library is established, and the training samples are expanded by selecting terrain factors, sample enhancement processing and image processing methods; a plurality of initial models with different neural network structures are constructed, and the plurality of initial models are trained by the training samples of the micro-terrain training sample library; through a plurality of rounds of dynamic optimization training strategies, the model parameters are adjusted and optimized, and the performance indicators corresponding to each model are determined; and the model satisfying the preset condition is selected as the micro-terrain identification model; Wherein, the micro-terrain training sample library is established, and the training samples are expanded by selecting terrain factors, sample enhancement processing and image processing methods, comprising: The following one or more ways are used for sample expansion: Selecting terrain factors, quantitatively analyzing the terrain factors, and retaining terrain factors meeting the preset requirements; Based on digital elevation model, digital orthophoto map and SHP format data, sample enhancement processing is performed; The training samples are preprocessed, and the preprocessing methods at least include projection, cropping and generating a gray scale image; Wherein, based on the digital elevation model, the digital orthophoto map and the SHP format data, the sample enhancement processing comprises: Taking the spatial coordinates of the grid center points of the digital elevation model as the characteristic variables and the elevation values as the dependent variables, a random forest regression model is constructed; by inserting a plurality of random points in the original grid and taking the spatial coordinates as the input, the elevation value is predicted by spatial interpolation to enhance the digital elevation model; the elevation mean difference between the grid center points and the random points inside the grid is calculated as an importance measurement index of the grid unit, and the grid center points with importance exceeding a preset importance threshold are selected and extracted as candidate terrain feature points; According to the candidate terrain feature points, a spatial distribution map of feature points and lines is drawn, and the digital elevation model under the original resolution is reconstructed; all grid points in the digital elevation model under the original resolution are taken as checkpoints to evaluate the error statistical indicators of the reconstructed digital elevation model and analyze the matching condition of the feature lines; According to the spatial distribution of the feature points, the error statistical indicators of the reconstruction and the matching condition of the feature lines, the sample enhancement processing is evaluated to determine the sample enhancement effect; Wherein, based on the digital elevation model, the digital orthophoto map and the SHP format data, the sample enhancement processing further comprises: The CART decision tree is used to divide the training sample space, and the minimum elevation deviation sum of squares of two sub-regions is taken as the division standard, and the minimum elevation deviation sum of squares is calculated by the following formula: where SS represents the minimum squared sum of elevation deviations of two sub-regions; z i is the elevation value of the sampling point in the region; is the average value of the elevation values of all sampling points in the region; Based on the divided training sample space, a plurality of training sample sets are generated by random sampling with replacement from the original terrain sampling point data; wherein the capacity of each training sample set remains the same as the original data set, the planar coordinates of the grid center point are taken as the input features, and the elevation is taken as the prediction target to train the CART decision tree model; the sample data is divided into leaf nodes through recursive partitioning of the planar coordinate space to form a complete regression tree; the regression tree is used as a simulation model of the terrain surface; By repeating the above sampling and tree building process, a plurality of regression trees are constructed to form a random forest model composed of a plurality of regression trees; wherein for random points in the same grid cell, the random forest provides a plurality of predicted elevation values, and the mean of the plurality of predicted elevation values is taken as the final elevation estimate of the corresponding random point. 2.The micro-terrain recognition method of fusing multi-source data and image processing technology according to claim 1, wherein, Select terrain factors, quantitatively analyze the terrain factors, and retain terrain factors that meet the preset requirements, including: According to the quantitative analysis of the variance inflation factor corresponding to each terrain factor, the variance inflation factor is calculated by the following formula: In the formula, is the square of the complex correlation coefficient of the jth variable to other variables; VIF j represents the VIF value corresponding to the jth variable, the larger the VIF value, the higher the degree of the variable being explained by other variables, and the stronger the redundancy; the variable with a VIF value greater than a preset threshold is removed, and the variable with a VIF value less than or equal to the preset threshold is retained. 3.The micro-terrain recognition method of fusing multi-source data and image processing technology according to claim 1, characterized in that, A micro-terrain training sample library is established, and the training samples are expanded by selecting terrain factors, sample enhancement processing and image processing methods, including: Based on stratified random sampling, a first number of training areas and a first number of test areas are selected; For the training area and the test area, the sample range is determined by fusing the preset resolution digital orthophoto map and using semi-automatic boundary extraction to assist manual interpretation to obtain the sample and label; wherein the samples of the training area and the test area are projected to the target coordinate system, and bicubic interpolation and resampling processing are performed to construct a digital orthophoto map containing RGB three bands and a feature data set of a plurality of terrain factors; For the digital orthophoto map and the feature data set of the terrain factor, the water body position and range are determined using a sliding window and SHP data to generate training samples; wherein after the training sample construction is completed, a part of the training images are subjected to data enhancement processing such as rotation or inversion. 4.The micro-terrain recognition method of fusing multi-source data and image processing technology according to claim 1, wherein, The training method of the micro-terrain recognition model includes: A basic backbone network with ResNet18 as the initial model is constructed, ACON is used as the activation function in the residual block of the initial model, a feature pyramid network is used as the neck of the initial model, and a convolutional layer is used as the head of the initial model. 5.The micro-terrain recognition method of fusing multi-source data and image processing technology according to claim 1, characterized in that, A plurality of initial models with different neural network structures are constructed, and the plurality of initial models are trained respectively by the training samples of the micro-terrain training sample library. After multiple rounds of dynamic optimization training strategy, the model parameters are adjusted and optimized, and the performance indicators corresponding to each model are determined. The model that meets the preset condition in the performance indicators is selected as the micro-terrain recognition model, including: Through traditional convolutional neural networks AlexNet, VGGNet, classic deep network ResNet, and introduction of attention mechanism ACON, RF+ResNet, ACON-ResNet, RF+ACON-ResNet, corresponding initial models are respectively constructed; After training and testing the multiple initial models respectively, quantitative analysis and qualitative evaluation are performed on each model to obtain corresponding performance indicators, and the model with the highest ranking of performance indicators is selected as the micro-terrain recognition model. 6.The micro-terrain recognition method of fusing multi-source data and image processing technology according to claim 5, characterized in that, The performance indicators at least include precision, recall, F1 score and prediction accuracy.

7. A micro-terrain recognition device that fuses multi-source data with image processing technology, characterized by, The device comprises: a data acquisition module configured to acquire multi-source data of a target region, wherein the multi-source data comprises remote sensing images, vector data and terrain factors; a micro-terrain recognition module configured to determine micro-terrain information of the target region by inputting the multi-source data of the target region as input samples into a micro-terrain recognition model; wherein the micro-terrain recognition module comprises: a sample processing unit configured to establish a micro-terrain training sample library, and expand training samples by selecting terrain factors, sample enhancement processing and image processing methods; a model training unit configured to construct multiple initial models with different neural network structures, train the multiple initial models by using training samples of the micro-terrain training sample library, adjust and optimize model parameters through multiple rounds of dynamic optimization training strategies, and determine performance indicators corresponding to each model; and select a model satisfying a preset condition as the micro-terrain recognition model; wherein the sample processing unit is specifically configured to: expand samples in one or more of the following ways: select terrain factors, quantitatively analyze the terrain factors, and retain terrain factors meeting preset requirements; perform sample enhancement processing based on digital elevation models, digital orthophoto maps and SHP format data; perform preprocessing on training samples, and the preprocessing methods at least include projection, cropping and generating grayscale images; wherein the sample enhancement processing based on digital elevation models, digital orthophoto maps and SHP format data comprises: construct a random forest regression model by taking spatial coordinates of grid center points of a digital elevation model as feature variables and taking elevation values as dependent variables; enhance the digital elevation model by inserting multiple random points in the original grid and predicting elevation values by spatial interpolation with spatial coordinates as input; calculate the mean difference of elevation between the grid center points and the random points inside the grid as an importance measurement index of the grid unit, and select grid center points with importance exceeding a preset importance threshold as candidate terrain feature points; draw a spatial distribution map of feature points and lines according to the candidate terrain feature points, and reconstruct the digital elevation model under the original resolution; take all grid points in the digital elevation model under the original resolution as checkpoints, evaluate error statistical indicators of the reconstructed digital elevation model, and analyze matching conditions of feature lines; evaluate the sample enhancement processing according to dimensions including spatial distribution of feature points, error statistical indicators of the reconstruction, and matching conditions of feature lines, and determine the sample enhancement effect; wherein the sample enhancement processing based on digital elevation models, digital orthophoto maps and SHP format data further comprises: divide the training sample space by using a CART decision tree, take the minimum elevation deviation sum of squares of two sub-regions as the division standard, and calculate the minimum elevation deviation sum of squares by using the following formula: where SS represents the minimum squared sum of elevation deviations of two sub-regions; z i is the elevation value of the sampling point in the region; is the average value of the elevation values of all sampling points in the region; Based on the divided training sample space, a plurality of training sample sets are generated by random sampling with replacement from the original terrain sampling point data; wherein the capacity of each training sample set remains the same as the original data set, the planar coordinates of the grid center point are taken as the input features, and the elevation is taken as the prediction target, and a CART decision tree model is trained; through recursive partitioning of the planar coordinate space, the sample data is divided into leaf nodes to form a complete regression tree; the regression tree is used as a simulation model of the terrain surface; By repeating the above sampling and tree building process, a plurality of regression trees are constructed to form a random forest model composed of a plurality of regression trees; wherein for a random point in the same grid cell, the random forest provides a plurality of predicted elevation values, and the mean of the plurality of predicted elevation values is taken as the final elevation estimate of the corresponding random point.

Images