Terrain key point selection method based on deep learning

Abstract

The invention relates to a terrain key point selection method based on deep learning, and the method comprises the steps: carrying out the preprocessing of battlefield environment data in a combat area, and obtaining a feature value matrix; inputting the characteristic value matrix into a pre-trained terrain key point selection model, and selecting to obtain terrain key points in the region; the terrain key point selection model selectively pays attention to and processes the most important part based on an attention mechanism, ignores irrelevant or secondary information, and then comprehensively measures interrelation of terrain elements in a battlefield environment through a convolutional neural network to realize selection of terrain key points. The terrain key point selection method designed by the invention is not only high in accuracy and simple in structure, but also has the advantages of good mobility, few parameters and low calculation complexity, and has relatively high practical application value.

Description

Technical Field

[0001] The present invention relates to a method for selecting key terrain points in a combat area, and in particular to a method for selecting key terrain points in a combat area based on an eigenvalue matrix, an attention mechanism, and a convolutional neural network. Background Art

[0002] Different morphological characteristics of terrain units give them varying basic combat capabilities. In certain combat contexts, their location and relationship to adjacent units also give them varying military value. Some terrain units possess both excellent basic combat capabilities and "enduring" military value. These terrain units, which play a crucial role in stabilizing defense zones, are often referred to as key terrain points and are among the most important targets to seize and control in modern warfare. To seize the initiative and determine the direction of a war, it is imperative to identify and control key terrain points within the combat zone, achieving the strategic objective of attacking without harming the enemy.

[0003] Analyzing key terrain points presents numerous ambiguities that must be resolved. Traditional methods, relying on commanders' operational experience and command skills, or relying on expert analysis during operational planning, are unable to meet the real-time and accurate decision-making requirements of modern warfare. To overcome the subjective influence of qualitative analysis and rapidly and accurately select and determine key terrain points, quantitative methods must be employed for battlefield environmental analysis. Neural networks can learn complex mappings between input and output data and express these mappings implicitly. Furthermore, in modern warfare, the volume of battlefield environmental information has increased significantly with the advancement of sensor, remote sensing, and communication technologies. Neural networks, with their powerful data processing and pattern recognition capabilities, demonstrate enormous potential for processing this large-scale, multi-source, and heterogeneous data.

[0004] The present invention uses the eigenvalue matrix to extract and process the terrain information in the combat area, and based on the attention mechanism and convolutional neural network, quantitatively mines the intrinsic connections within the terrain unit and between adjacent units, and finally realizes the rapid selection of terrain key points. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for selecting key terrain points in a combat area, which assists commanders in quickly and accurately selecting key terrain points in the combat area so as to complete strategic deployment, execute tactical actions, and grasp the battlefield situation.

[0006] The present invention uses the eigenvalue matrix to extract element information within the combat area, adds an attention module to make the model pay more attention to places with greater influence, thereby improving the accuracy of judgment, and finally analyzes the relationship between adjacent terrain units through a convolutional neural network, thereby comprehensively considering the interactive information between terrain units and finally outputting it.

[0007] To achieve the above objectives, the present invention provides a method for selecting key terrain points based on deep learning, comprising the following steps:

[0008] The battlefield environment data in the combat area is preprocessed to obtain the eigenvalue matrix.

[0009] The eigenvalue matrix is input into the pre-trained terrain key point selection model to select the terrain key points in the area.

[0010] The terrain key point selection model selectively focuses on and processes the most important parts based on the attention mechanism, ignoring irrelevant or minor information, and then uses a convolutional neural network to comprehensively measure the interrelationships between terrain elements in the battlefield environment to achieve the selection of terrain key points.

[0011] The method also includes a training step for a terrain key point selection model, which is achieved by using labeled battlefield environment data to establish a training set, inputting sample data in the training set into the battlefield environment data in batches, calculating the cross entropy loss function, performing gradient backpropagation, and using the Adam optimizer to update the network weights. The update is iteratively updated until the training requirements are met to obtain a trained terrain key point selection model.

[0012] The method includes the following steps for preprocessing battlefield environment data within the combat area:

[0013] Step 1: Discrete the combat area space according to its spatial range and spatial resolution.

[0014] Step 2: Quantify the environmental elements of the combat area based on grid units.

[0015] Extract data from the battlefield environment database, including height, slope, surface fragmentation, residential area level, vegetation level, water system level, and road level, and assign values to each grid:

[0016] gd i ={h i ,s i ,sf i ,b i ,v i ,w i ,r i}

[0017] where gd i represents the i-th grid, h i ,s i ,sf i ,b i ,v i ,w i ,ri They represent the height, slope, surface fragmentation, residential area level, vegetation level, water system level, and road level data of the i-th grid respectively.

[0018] Step 3: Normalize the data so that its range is between 0 and 1, thereby obtaining the eigenvalue matrix.

[0019] Preferably, the input of the terrain key point selection model described in the above method is the eigenvalue matrix, and the output is the extraction result of whether any grid point in the area is a terrain key point. The grid output result of 0 indicates that the grid is not a terrain key point, and the grid output result of 1 indicates that the grid is a terrain key point.

[0020] Preferably, the terrain key point selection model described in the above method includes a CBAM module, a CNN module, an MLP module, and a Softmax layer in sequence.

[0021] Preferably, the CBAM module described in the above method includes a channel attention module and a spatial attention module. The input of the CBAM module is an eigenvalue matrix, which is multiplied element-by-element with the output features of the channel attention module and the spatial attention module in sequence, and the output is an attention-enhanced feature matrix.

[0022] Compared with the prior art, the advantages of the present invention are:

[0023] 1. This paper innovatively designs a terrain key point selection model based on a convolutional neural network with an attention mechanism. The attention mechanism module can automatically learn the correlation between different grids of the eigenvalue matrix, thereby capturing richer environmental information. By extracting local features through the convolutional neural network, it can quickly and accurately select terrain key points within the combat area.

[0024] 2. The terrain key point selection model of the present invention greatly reduces the number of parameters that need to be learned in the network by applying convolutional neural networks, so that the model can quickly converge to the optimal solution. The model has high generalization and high practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of model training and prediction of the present invention;

[0026] Figure 2 The model structure diagram is selected for the terrain key points of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0029] This method targets seven aspects of data extracted from the battlefield environment system within the combat area. The specific forms are as follows:

[0030] Height: the specific value of the local elevation;

[0031] Slope: the specific value of the local slope;

[0032] Surface fragmentation: the specific value of the local surface fragmentation;

[0033] Settlement level: Levels classified by military application;

[0034] Vegetation Class: Classification based on military applications;

[0035] Water system level: Level divided by military application;

[0036] Road Grade: Grades classified by military applications.

[0037] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0038] like Figure 1 As shown, an embodiment of the present invention provides a method for selecting terrain key points based on deep learning, which specifically includes the following steps:

[0039] Step 1: The first step is data preprocessing:

[0040] 1) Discretely divide the combat area space according to its spatial range and spatial resolution;

[0041] 2) Quantify the environmental elements of the combat area based on grid units;

[0042] Extract data on seven aspects, including height, slope, surface fragmentation, residential area level, vegetation level, water system level, and road level, from the battlefield environment database and assign values to each grid:

[0043] gd i ={h i ,s i ,sf i ,b i ,v i,w i ,r i}

[0044] where gd i represents the i-th grid, h i ,s i ,sf i ,b i ,v i ,w i ,r i They represent seven types of data, namely, the height, slope, surface fragmentation, residential area level, vegetation level, water system level, and road level of the i-th grid.

[0045] 3) Normalize the data so that it ranges from 0 to 1 to obtain the eigenvalue matrix. The specific formula for each grid is as follows:

[0046]

[0047] Among them, x i In turn, {h i ,s i ,sf i ,b i ,v i ,w i ,r i}, min(x i ) means that each x i The minimum value, min(x i ) means that each x i The maximum value of x i ′ Represents each x i Normalized data.

[0048] Step 2: The second step is to build a neural network model. The terrain key point selection model structure designed by the present invention is as follows Figure 2 As shown in Figure 2, the selection model mainly includes the following components:

[0049] 1) CBAM module: This module includes a channel attention module and a spatial attention module. The input of this module is the eigenvalue matrix obtained by preprocessing, and the output is the attention-enhanced feature matrix. The calculation formula of the channel attention module is:

[0050] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0051] Where F is the input, AvgPool and MaxPool represent global average pooling and maximum pooling operations respectively, MLP represents multi-layer perceptron, and σ represents the Sigmoid activation function.

[0052] The calculation formula of the spatial attention module is:

[0053] M s (F)=σ(f([AvgPool(F);MaxPool(F)]))

[0054] Among them, F is the input, f represents the convolution operation, [AvgPool(F); MaxPool(F)] means splicing the average pooling and maximum pooling results along the channel axis, and σ represents the Sigmoid activation function.

[0055] 2) CNN module: The attention-enhanced feature matrix calculated by the CBAM module is input into the 3*3 standard convolution layer and the average pooling layer in sequence, and finally passes through the ReLU function to obtain the output result.

[0056] 3) MLP module: The results of the CNN module are sequentially passed through the fully connected layer and the nonlinear activation function to obtain the output result.

[0057] 4) Softmax layer: The output of the MLP module is output through the Softmax function. The specific calculation formula is:

[0058]

[0059] Among them, z i Represents each result output by the MLP module, and n represents the number of elements.

[0060] When the calculation result is less than 0.5, the output is 0; when the calculation result is greater than 0.5, the output is 1.

[0061] Step 2: After the model is built, it is trained and evaluated using the existing training and validation sets. The cross-entropy loss function is used during training, and the Adam optimizer is used as the optimizer. During the training phase, the model is trained using the training set. Sample data is fed into the model in batches, the cross-entropy loss function is calculated, gradient backpropagation is performed, and the network weights are updated using the Adam optimizer. The training is then terminated after it iterates until the accuracy on the validation set no longer improves.

[0062] Step 3: After the training is completed, in order to test the effect of the terrain key point selection model designed by the present invention, we analyze the model test results on the test dataset and conduct a comparative experiment with the existing change detection method. The accuracy evaluation indicator used is the detection accuracy.

[0063] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and are intended to be encompassed by the claims of the present invention.

Claims

1. A method for selecting terrain key points based on deep learning, characterized in that: include: Preprocess the battlefield environment data within the combat area to obtain the eigenvalue matrix; Input the eigenvalue matrix into the pre-trained terrain key point selection model to select the terrain key points in the area; The terrain key point selection model selectively focuses on and processes the most important parts based on the attention mechanism, ignoring irrelevant or minor information. It then uses a convolutional neural network to comprehensively measure the interconnectedness of terrain elements in the battlefield environment to achieve the selection of terrain key points. The method also includes a step of training a terrain key point selection model, which is to establish a training set by using labeled battlefield environment data, input sample data in the training set into the battlefield environment data in batches, calculate the cross entropy loss function, perform gradient backpropagation, and use the Adam optimizer to update the network weights. Iterative updates are performed until the training requirements are met, thereby obtaining a trained terrain key point selection model; The method of preprocessing battlefield environment data within the combat area specifically includes the following steps: Step 1: Discrete the combat area space according to the spatial range and spatial resolution of the combat area; Step 2: Quantify the environmental elements of the combat area based on grid units; Extract data from the battlefield environment database, including height, slope, surface fragmentation, residential area level, vegetation level, water system level, and road level, and assign values to each grid: gd i ={h i ,s i ,sf i ,b i ,v i ,w i ,r i } where gd i represents the i-th grid, h i ,s i ,sf i ,b i ,v i ,w i ,r i Respectively represent the height, slope, surface fragmentation, residential area level, vegetation level, water system level, and road level data of the i-th grid; Step 3: Normalize the data so that its range is between 0 and 1, thereby obtaining the eigenvalue matrix.

2. The method for selecting terrain key points based on deep learning according to claim 1, characterized in that: The input of the terrain key point selection model is the eigenvalue matrix, and the output is the extraction result of whether any grid point in the area is a terrain key point. The grid output result of 0 indicates that the grid is not a terrain key point, and the grid output result of 1 indicates that the grid is a terrain key point.

3. The method for selecting terrain key points based on deep learning according to claim 1, characterized in that: The terrain key point selection model includes a CBAM module, a CNN module, an MLP module, and a Softmax layer in sequence.

4. The method for selecting terrain key points based on deep learning according to claim 3, characterized in that: The CBAM module includes a channel attention module and a spatial attention module. The input of the CBAM module is an eigenvalue matrix, which is multiplied element by element with the output features of the channel attention module and the spatial attention module in sequence, and the output is an attention-enhanced feature matrix.

Images