Saddle point extraction method, device, equipment, storage medium and program product

Abstract

The application relates to the technical field of topographic analysis, in particular to a saddle point extraction method, device, equipment, storage medium and program product, wherein the method comprises the following steps: acquiring DEM data for topographic saddle extraction; based on a preset positive and negative topography method, extracting a candidate saddle point set from the DEM data, and extracting a plurality of topographic features of each candidate saddle point in the candidate saddle point set at a plurality of scales; inputting the plurality of topographic features of each candidate saddle point at the plurality of scales into a pre-established convolutional neural network for multi-scale topographic feature fusion, and removing false saddle points in the candidate saddle point set according to an output result to obtain a final saddle point. Therefore, by combining the positive and negative topography method and the multi-scale feature fusion convolutional neural network, the problems of excessive false saddle points when the positive and negative topography method extracts saddle points are solved, the extraction precision and accuracy of the saddle points are improved, and intelligent analysis and identification of topography are promoted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of topographic analysis, and in particular to a saddle point extraction method, device, equipment, storage medium and program product. BACKGROUND

[0002] As one of the terrain features, the saddle has important significance in the fields of terrain feature extraction and terrain modeling. The saddle is located at the intersection of the lowest part of the ridge line and the highest part of the valley bottom line. The accurate extraction of the saddle has always been a hot and difficult topic. The existing extraction methods include mathematical definition method, contour method and hydrological method. Among them, the positive-negative terrain method (PNTM) is widely used.

[0003] In related technologies, to ensure the continuity of the valley bottom line, the positive-negative terrain method is inspired by the continuity of the ridge line. The original terrain is inverted into an inverse terrain. The continuous ridge line of the inverse terrain is the continuous mountain control line of the positive terrain, which avoids the situation that the discontinuous mountain control line of the original terrain leads to insufficient extraction of the saddle.

[0004] However, due to the complexity of the terrain, the highest part of the valley bottom line is usually disconnected, and there is no intersection with the ridge line, which easily causes the loss of saddle points. At the same time, this method extracts a large number of false saddle points. At present, the false saddle points are mainly removed by manual judgment. The traditional physical mechanism model is mainly used to extract terrain features based on elevation data, without considering the multi-scale effect problem of complex terrain, which needs to be solved urgently. SUMMARY

[0005] The present application provides a saddle point extraction method, device, equipment, storage medium and program product to solve the problem of excessive false saddle points when extracting saddle points by the positive-negative terrain method, improve the extraction accuracy and accuracy of saddle points, and promote the intelligent analysis and recognition of topography.

[0006] The first aspect embodiment of the present application provides a saddle point extraction method, including the following steps:

[0007] Obtain DEM data for terrain saddle extraction;

[0008] Based on the preset positive-negative terrain method, extract a candidate saddle point set from the DEM data, and extract a plurality of terrain features of each candidate saddle point in the candidate saddle point set at a plurality of scales;

[0009] Input the plurality of terrain features of each candidate saddle point at a plurality of scales into a pre-established convolutional neural network for multi-scale terrain feature fusion, and remove false saddle points in the candidate saddle point set according to the output result to obtain a final saddle point.

[0010] Optionally, the preset positive and negative terrain method is used to extract a candidate saddle point set from the DEM data, comprising:

[0011] A reverse DEM data is determined according to the DEM data, and a ridge line is extracted based on the DEM data, and a mountain control line is extracted based on the reverse DEM data;

[0012] An intersection data of the ridge line and the mountain control line is determined, and the intersection data is preprocessed to obtain the candidate saddle point set.

[0013] Optionally, before inputting the terrain features of each candidate saddle point at multiple scales into the pre-established multi-scale terrain feature fusion convolutional neural network, the method further comprises:

[0014] A target saddle point set for training is obtained;

[0015] A plurality of terrain parameters in a preset area are extracted with each saddle point in the target saddle point set as a center to obtain terrain features of each saddle point in the target saddle point set;

[0016] A plurality of saddle control areas of multiple scales are determined, and a preset activation function, a preset loss function and the plurality of saddle control areas of multiple scales are used to train the preset convolutional neural network with the terrain features of each saddle point in the target saddle point set to obtain the pre-established multi-scale terrain feature fusion convolutional neural network.

[0017] Optionally, the preset activation function is:

[0018]

[0019] wherein y i is an output value after the activation function, x i is an output value after weight multiplication at each level, and a is a constant between 0 and 1.

[0020] Optionally, the preset loss function is:

[0021]

[0022] wherein Loss is a total loss function, N is a sample number, K is a classification number, x ij is an output value of a full connection layer, i is an i-th sample, j is a j-th result (including two results of saddle and pseudo-saddle), f(x ij ) is a probability after normalization of a softmax activation function, and y ij is a label probability.

[0023] The second aspect embodiment of the application provides a saddle point extraction device, comprising:

[0024] An acquisition module is configured to acquire DEM data for terrain saddle extraction.

[0025] An extraction module is configured to extract a candidate saddle point set from the DEM data based on a preset positive-negative terrain method, and extract a plurality of terrain features of each candidate saddle point in the candidate saddle point set at a plurality of scales.

[0026] A determination module is configured to input the plurality of terrain features of each candidate saddle point at a plurality of scales into a pre-established convolutional neural network for multi-scale terrain feature fusion, and remove false saddle points in the candidate saddle point set according to an output result to obtain a final saddle point.

[0027] Optionally, the extraction module is specifically configured to:

[0028] determine inverse DEM data according to the DEM data, extract a ridge line based on the DEM data, and extract a mountain control line based on the inverse DEM data;

[0029] determine intersection data of the ridge line and the mountain control line, and pre-process the intersection data to obtain the candidate saddle point set.

[0030] Optionally, the determination module is further configured to:

[0031] acquire a target saddle point set for training;

[0032] extract a plurality of terrain parameters in a preset area with each saddle point in the target saddle point set as a center to obtain terrain features of each saddle point in the target saddle point set;

[0033] determine a plurality of scales of saddle control areas, and train a preset convolutional neural network based on a preset activation function, a preset loss function, and the plurality of scales of saddle control areas, and the terrain features of each saddle point in the target saddle point set to obtain the pre-established convolutional neural network for multi-scale terrain feature fusion.

[0034] Optionally, the preset activation function is:

[0035]

[0036] wherein y i is an output value after the activation function, x i is an output value after weight multiplication at each level, and a is a constant between 0 and 1.

[0037] Optionally, the preset loss function is:

[0038]

[0039] wherein, Loss is a total loss function, N is a sample number, K is a classification number, x ij is a value of a full connection layer output, i is an i-th sample, j is a j-th class result (including two class results of a saddle or pseudo saddle), f(x ij ) is a probability after a softmax activation function normalization, y ij is a label probability.

[0040] The third aspect of the application provides an electronic device, comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to execute the saddle point extraction method as described in the above embodiments.

[0041] The fourth aspect of the application provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the saddle point extraction method as described in the above embodiments.

[0042] The fifth aspect of the application provides a computer program product, which stores a computer program, and the program is executed by a processor to implement the saddle point extraction method as described in the above embodiments.

[0043] Therefore, after obtaining the DEM data for terrain saddle extraction, the candidate saddle point set is extracted from the DEM data based on the preset positive and negative terrain method, and the multiple terrain features of each candidate saddle point in the candidate saddle point set at multiple scales are extracted, the multiple terrain features of each candidate saddle point at multiple scales are input into the convolutional neural network for multi-scale terrain feature fusion which is established in advance, and the pseudo saddle points in the candidate saddle point set are removed according to the output result to obtain the final saddle point. Therefore, by combining the positive and negative terrain method and the multi-scale feature fusion convolutional neural network, the problem of excessive pseudo saddle points when the positive and negative terrain method extracts the saddle point is solved, the extraction precision and accuracy of the saddle point are improved, and the intelligent analysis and identification of the terrain and landscape are promoted.

[0044] Additional aspects and advantages of the application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0045] The above and / or additional aspects and advantages of the application will become apparent and be readily appreciated from the following description, including the appended drawings.

[0046] Figure 1A flowchart of a saddle point extraction method according to an embodiment of the present application is provided.

[0047] Figure 2 A flowchart of a saddle point extraction method according to an embodiment of the present application is provided.

[0048] Figure 3 A flowchart of a saddle point extraction method according to an embodiment of the present application is provided.

[0049] Figure 4 A flowchart of a saddle point extraction method according to an embodiment of the present application is provided.

[0050] Figure 5 A flowchart of a saddle point extraction method according to an embodiment of the present application is provided.

[0051] Figure 6 A flowchart of a saddle point extraction method according to an embodiment of the present application is provided.

[0052] Figure 7 A flowchart of a saddle point extraction method according to an embodiment of the present application is provided.

[0053] Figure 8 A flowchart of a saddle point extraction method according to an embodiment of the present application is provided. DETAILED DESCRIPTION

[0054] Embodiments of the present application are described in detail below with reference to the accompanying drawings. The same or similar components are denoted by the same or similar reference numerals throughout the drawings, and repeated description thereof is omitted. The embodiments described below are examples for explaining the present application, and are not intended to be limiting of the present application.

[0055] Before introducing the saddle point extraction method of the embodiments of the present application, several saddle point extraction methods in the related art are briefly introduced.

[0056] The mathematical definition method uses the idea of moving window to traverse and calculate two characteristic images of each pixel, i.e., the adjacent elevation difference and the rate of change of the elevation difference, and finally determines the pixel to which the saddle belongs based on the positive and negative differences of the elevation difference in the orthogonal direction. The method of continuous derivation is further used to obtain the second derivative and the curvature direction of the adjacent pixel, and the saddle pixel is identified by combining the positive and negative of the second derivative and the orthogonality of the curvature direction. Since the mathematical definition method needs to traverse each pixel and the window setting has uncertainty, the spatial topological relationship between the saddle and the surrounding terrain is easily ignored, and the terrain elements of the saddle are not considered, and a large amount of computing resources is consumed.

[0057] The contour method uses the spatial topological relationship of "mountain top point-saddle point-mountain top point" to extract the saddle point. The saddle is located in the low-lying part between two adjacent mountain tops, and in the contour map, it is represented as the middle part of two parallel positive contour clusters contained in a mountain body. Based on this characteristic, the closed contour cluster is converted into an elevation surface data according to certain rules to construct an elevation surface tree, and then the adjacent topological relationship between the elevation surfaces is used to realize recursive searching and extraction of the saddle. This method ensures the equal relationship and spatial proximity relationship between the number of saddles and the number of mountain tops, and ensures the extraction accuracy of the saddle. However, due to the influence of complex terrain, this method can only be used for experiments in a small area with simple and closed contour cluster topological relationship, and has not been verified in complex large area saddle extraction experiments.

[0058] The hydrological analysis method uses the principle of hydrology to extract terrain characteristic lines, and takes the intersection point of the terrain characteristic lines as the saddle point. Since the saddle is located at the intersection of the lowest part of the ridge line and the highest part of the valley bottom line, the water flow division line is extracted as the ridge line and the water flow collection line as the valley bottom line (also known as the mountain control line), and the intersection of the two is extracted by using the topological intersection method, which is the saddle. The positive and negative terrain method is a representative method of the hydrological analysis method, and is also a commonly used example in current textbooks. In order to ensure the continuity of the valley bottom line, the method is inspired by the continuity of the ridge line, and the original terrain is inverted into an inverse terrain. The continuous ridge line of the inverse terrain is extracted, which is the continuous mountain control line of the positive terrain, avoiding the situation of insufficient saddle extraction due to the broken mountain control line of the original terrain. However, this method uses the method of dividing the water line and the positive terrain to eliminate the error influence of the negative terrain such as the depression, and inevitably extracts part of the broken ridge line, so there are still a large number of missed and misjudged saddles.

[0059] Since the mathematical definition method lacks consideration of the terrain background and the spatial relationship of the saddle, the contour method is easily constrained by data quality and algorithm complexity, and the hydrological analysis method is the most commonly used method for saddle point extraction. Because the hydrological elements are the most direct results of the terrain undulation, combining hydrological knowledge to extract the saddle can improve the accuracy of saddle extraction.

[0060] Based on the above problems, the saddle point extraction method is proposed. In the method, DEM data for terrain saddle extraction is obtained; based on the preset positive and negative terrain method, a candidate saddle point set is extracted from the DEM data, and a plurality of terrain features of each candidate saddle point in the candidate saddle point set at a plurality of scales are extracted; the plurality of terrain features of each candidate saddle point at a plurality of scales are input into a pre-established multi-scale terrain feature fusion convolutional neural network, and false saddle points in the candidate saddle point set are removed according to the output result to obtain a final saddle point. Therefore, by combining the positive and negative terrain method and the multi-scale feature fusion convolutional neural network, the problem of excessive false saddle points when the positive and negative terrain method extracts the saddle point is solved, the extraction accuracy and accuracy of the saddle point are improved, and the intelligent analysis and identification of the terrain and landscape are promoted.

[0061] Specifically, Figure 1 A flowchart of a saddle point extraction method provided by an embodiment of the application is shown.

[0062] As Figure 1 shown, the saddle point extraction method includes the following steps:

[0063] In step S101, DEM data for terrain saddle extraction is obtained.

[0064] The saddle point is a low-lying part between two adjacent mountain tops in the shape of a saddle, and is the intersection of the lowest part of the ridge and the highest part of the valley, that is, the point indicating the spatial position of the saddle is called the saddle point. DEM (Digital Elevation Model) is a data form representing elevation information, which is a data set describing the shape of the earth's surface through a series of regularly or irregularly distributed points.

[0065] Specifically, for the extraction of terrain saddle points, LIDAR data, regular grid DEM data and regular triangulated irregular network (TIN) DEM data are usually selected, and the regular grid DEM data is the most commonly used terrain saddle extraction data. After obtaining the DEM data, the data needs to be preprocessed, including but not limited to data format conversion, splicing of different DEM pieces, noise removal, error correction, etc., to ensure the quality of the DEM data and meet the needs of the saddle point extraction task. In order to facilitate the analysis and processing of DEM data, it is usually necessary to import it into a geographic information system (GIS) software or other professional terrain analysis tool to better view the DEM data and perform necessary processing steps.

[0066] In step S102, a candidate saddle point set is extracted from the DEM data based on a preset positive-negative terrain method, and a plurality of terrain features of each candidate saddle point in the candidate saddle point set are extracted at a plurality of scales.

[0067] The preset positive-negative terrain method is a technique for extracting a terrain saddle point from DEM data, which uses a combination of positive and negative DEM data to determine a saddle point by extracting a ridge line and a mountain control line (i.e., a valley bottom line); and the candidate saddle point set refers to potential saddle points preliminarily screened by the positive-negative terrain method.

[0068] Specifically, the negative DEM data is determined based on the original DEM data, which is obtained by taking the elevation values in the original DEM data as negative. This is done to extract the mountain control line in the subsequent steps. The ridge line is extracted based on the DEM data using a hydrological analysis method. The ridge line is the dividing line of water flow in the terrain and can be determined by calculating the flow direction and convergence. The mountain control line is extracted based on the negative DEM data. The mountain control line reflects the valley bottom line in the terrain and can also be extracted by the hydrological analysis method. The intersection data of the ridge line and the mountain control line is found, and these intersection points are the candidate saddle points. The intersection data is preprocessed to obtain the candidate saddle point set. To extract a plurality of terrain features of each candidate saddle point in the candidate saddle point set at a plurality of scales, different terrain analysis scales need to be selected, which are defined as terrain features at different scales. At each selected scale, a plurality of terrain features of the candidate saddle point are extracted.

[0069] Alternatively, in some embodiments, the candidate saddle point set is extracted from the DEM data based on the preset positive-negative terrain method, which includes determining the negative DEM data based on the DEM data, extracting the ridge line based on the DEM data, and extracting the mountain control line based on the negative DEM data; determining the intersection data of the ridge line and the mountain control line, and preprocessing the intersection data to obtain the candidate saddle point set.

[0070] It can be understood that by combining the positive and negative DEM data, the ups and downs of the terrain can be more comprehensively captured, thereby preliminarily screening out regions that may contain saddle points. The positive DEM data is used to extract the ridge line, reflecting the convex features of the terrain; the negative DEM data is used to extract the mountain control line, reflecting the concave features of the terrain; the combination of the positive-negative terrain method can avoid the insufficient extraction of saddle points due to the complexity of the terrain, and ensure that as many potential saddle points as possible are extracted; although the preliminary screening may contain some false saddle points, through the further identification of the subsequent neural network, these misjudgment points can be effectively reduced, and the accuracy of the saddle point extraction can be improved.

[0071] Further, the saddle point extraction method is described in detail in combination with an embodiment of the present application, and the steps are as follows:

[0072] The saddle is located at the lower concave connection between two mountain tops. The topography at this location requires a protrusion in one direction and a depression in the other direction, where the protrusion boundary coincides with the mountain control line and the depression boundary coincides with the ridge line (as shown in Figure 3 Therefore, the positive-negative topography method is used to extract the candidate saddle points. First, the ridge line is extracted based on the DEM data, then the mountain control line is extracted based on the inverse DEM data, and finally the intersection of the ridge line and the mountain control line is obtained after preprocessing to obtain the candidate saddle points.

[0073] Step 1: Extract the ridge line Rl. This step extracts the ridge line by hydrological analysis and uses the positive terrain Tpos to eliminate the influence of negative terrain Tneg.

[0074] Step 1.1: Calculate the mean DEM, denoted as The expression is:

[0075]

[0076] where DEM is the original elevation data, and Neighborhoodmeananalysis(·) is the mean neighborhood calculation with a 11x11 window.

[0077] Calculate the difference between the original elevation data DEM and , where the difference greater than 0 is the positive terrain Tpos, and the difference less than 0 is the negative terrain, which can be expressed as:

[0078]

[0079]

[0080] Step 1.2: Extract the watershed boundary line based on the D8 algorithm. First, set a small elevation difference h, and perform filling processing on the DEM to eliminate the influence of ambiguous terrain, obtaining the DEM filled ; second, according to the elevation difference relationship, calculate the flow direction relationship of each grid of DEM, obtain the flow direction grid Flow_dir, which can be expressed as:

[0081] Flow_dir = Flowdirection(DEM filled )

[0082] Then, based on the flow direction grid, calculate the flow accumulation grid Flow_acc, which can be expressed as:

[0083] Flow_acc = Flowaccumulation(Flow_dir)

[0084] Finally, the grid cells with flow accumulation of 0 are extracted, which are usually the basin boundaries, and are denoted as Wl. The flow direction and flow accumulation calculations above can be completed based on ArcGIS tools.

[0085] Step 1.3: Extract the ridge line Rl excluding the negative terrain effect, which can be expressed as:

[0086] Rl = Wl x Ter pos

[0087] Step 2: Extract the mountain control line Sl. This step refers to Step 1 to extract the mountain control line based on the inverse DEM data, and uses the negative terrain Tneg to eliminate the effect of the positive terrain Tpos.

[0088] Step 2.1: Calculate the inverse DEM data which can be expressed as:

[0089]

[0090] Step 2.2: Based on extract the mountain control line using the D8 algorithm. This includes pore filling processing, calculating flow direction grid confluence grid extracting the inverse basin boundary The above process can be expressed as:

[0091]

[0092]

[0093]

[0094] wherein is the inverse DEM after pore filling processing.

[0095] Step 2.3: Extract the mountain control line Sl excluding the positive terrain effect, which can be expressed as:

[0096]

[0097] Step 3: Extract the candidate saddle point set S. This set is the intersection of the ridge line Rl and the mountain control line Sl, which can be expressed as:

[0098] S = Rl ∩ Sl

[0099] The saddle points obtained at this time are shown in Figure 4 (b), (c) and (d), and one saddle point corresponds to multiple grid cells, which can easily lead to a large number of repeated cases in the vectorization process. Therefore, we perform dimensionality reduction preprocessing for the case where one saddle point corresponds to one grid block, as shown in Figure 4(b1), (c1) and (d1) shown, taking any one grid of the grid block represents the saddle point. The pre-processing operation to obtain a candidate set of saddle points S.

[0100] In step S103, the multiple terrain features of each candidate saddle point at multiple scales are input into a pre-established multi-scale terrain feature fusion convolutional neural network, and false saddle points in the candidate saddle point set are removed according to the output result to obtain the final saddle point.

[0101] Wherein, the multi-scale feature fusion convolutional neural network (Convolutional Neural Network, CNN) is a convolutional neural network that accepts different scale data input through multiple channels and provides feature fusion.

[0102] Specifically, a convolutional neural network for saddle decision containing parallel extraction of terrain information, multi-channel terrain information fusion and multi-scale terrain information fusion is built; the network includes parallel extraction of terrain information, five independent parallel convolutional neural networks are set for each scale at different scales, a total of 20 parallel convolutional neural networks are used for feature extraction; multi-channel terrain information fusion, deep terrain semantic information of each channel is combined and convolution operation is performed; multi-scale terrain information fusion for saddle decision includes three hidden layers, terrain information of each scale is input into three fully connected layers through channel splicing and flattening, and the judgment result of the saddle point is output; the activation function is designed, the Leaky Relu function is used as the activation function to improve the nonlinear modeling ability of the network; the loss function is designed, the binary cross entropy is used as the loss function to quantify the difference between the training result and the actual label result, and adjust the weights in the model; the multi-scale terrain features of each candidate saddle point are input into the pre-trained convolutional neural network, and whether the saddle point is retained is judged according to the output result of the network; if the network output indicates that the point is a real saddle point, the point is retained; if it is a false saddle point, the point is removed; through the above steps, the false saddle points in the candidate saddle point set are removed, and the final obtained will be a verified real saddle point set.

[0103] Optionally, in some embodiments, the pre-set activation function is:

[0104]

[0105] Wherein, y i is the output value after the activation function, x i is the output value after the weight multiplication of each level, and a is a constant between 0 and 1.

[0106] It can be understood that the activation function ensures that all negative values have a non-zero slope, avoiding the problem of early gradient vanishing or gradient explosion in the training process, and helping to stabilize the training of the neural network; when constructing the multi-scale terrain feature fusion convolutional neural network, the activation function is applied to each level to ensure that the network can learn the complex nonlinear relationship of the saddle point extraction; by introducing the activation function, the neural network can better simulate the decision-making behavior of experts, thereby improving the identification efficiency of the pseudo-saddle point.

[0107] Optionally, in some embodiments, the preset loss function is:

[0108]

[0109] wherein Loss is the total loss function, N is the number of samples, K is the number of classifications, x ij is the value output by the full connection layer, i is the i-th sample, j is the j-th class result (here including the saddle or pseudo-saddle two class results), f(x ij ) is the normalized probability of the softmax activation function, y ij is the label probability.

[0110] It can be understood that the loss function is mainly used in the model training stage to quantify the difference between the training result and the actual label result, and adjust the weights in the model based on the loss; by calculating the difference between the actual label probability and the model prediction probability, the performance of the model is evaluated. In the training process, the model will continuously adjust its parameters to minimize the loss function, thereby improving the extraction accuracy and accuracy of the model for the saddle point; combined with the multi-scale feature fusion convolutional neural network, the method can fully consider different scale terrain features when extracting the saddle point, thereby improving the accuracy and robustness of the extraction.

[0111] Therefore, after obtaining the DEM data for terrain saddle extraction, based on the preset positive and negative terrain method, the candidate saddle point set is extracted from the DEM data, and the multiple terrain features of each candidate saddle point in multiple scales are extracted. Each candidate saddle point in multiple scales is input into the pre-established multi-scale terrain feature fusion convolutional neural network, and the pseudo-saddle points in the candidate saddle point set are removed according to the output result, and the final saddle point is obtained. Therefore, by combining the positive and negative terrain method and the multi-scale feature fusion convolutional neural network, the problem of excessive pseudo-saddle points when extracting the saddle point by the positive and negative terrain method is solved, the extraction accuracy and accuracy of the saddle point are improved, and the intelligent analysis and identification of the terrain and landscape are promoted.

[0112] Optionally, in some embodiments, before inputting the plurality of terrain features of each candidate saddle point at a plurality of scales to the pre-established multi-scale terrain feature fusion convolutional neural network, further comprising: obtaining a target saddle point set for training; extracting a plurality of terrain parameters in a preset area centered on each saddle point in the target saddle point set to obtain terrain features of each saddle point in the target saddle point set; determining a plurality of scale saddle control areas, and training the preset convolutional neural network based on the preset activation function, the preset loss function and the plurality of scale saddle control areas using the terrain features of each saddle point in the target saddle point set to obtain the pre-established multi-scale terrain feature fusion convolutional neural network.

[0113] wherein the target saddle point set refers to the final real saddle point set determined after processing by the multi-scale terrain feature fusion convolutional neural network.

[0114] Further, an embodiment of the present application is combined to explain in detail how to establish a multi-scale terrain feature fusion convolutional neural network, and the steps are as follows:

[0115] The multi-scale terrain feature fusion convolutional neural network generally includes three contents: terrain feature extraction, multi-scale terrain feature construction, and multi-scale terrain feature fusion convolutional neural network construction.

[0116] Step 1: terrain feature extraction. Here, five terrain parameters of elevation E, aspect A, curvature C, slope S and hillshade H in a specific area centered on each saddle point are extracted as input parameters of the convolutional neural network, each parameter is a gray image, and referring to the three channels of the RGB color image, each gray image is regarded as a channel, and the input image of the convolutional neural network is a multi-channel fusion image. It can be expressed as:

[0117] E = Elevation (DEM)

[0118] A = Aspect (DEM)

[0119] C = Curve (DEM)

[0120] S = Slope (DEM)

[0121] H = Hillshade (DEM)

[0122] Step 2: Constructing multi-scale terrain features. Here, the area of the same level surface at the saddle elevation of the saddle and the two peaks is called the saddle control area. Because the distance, elevation and size of the mountain of different saddles are different, the saddle control area is also different, so the spatial range of the saddle range extraction reference also exists difference. In order to make the terrain information fed into the convolutional neural network cover all the control area of the saddle, and can combine different scale terrain features to make decisions, here four scales of saddle control area are set, and the control area window size corresponding to each scale is the smallest integer value, the largest integer value and two four equal values of the saddle control area of all samples, respectively 400m*400m, 800m*800m, 1200m*1200m, 1600m*1600m, as shown in Figure 5 . The terrain features of each scale are denoted as Feature 400 ={E 400 ,A 400 ,C 400 ,S 400 ,H 400}, Feature 800 ={E 800 ,A 800 ,C 800 ,S 800 ,H 800}, Feature 1200 ={E 1200 ,A 1200 ,C 1200 ,S 1200 ,H 1200}, Feature 1600 ={E 1600 ,A 1600 ,C 1600 ,S 1600 ,H 1600}.

[0123] Step 3: Establishing a convolutional neural network for multi-scale terrain feature fusion. The network includes parallel extraction of terrain information, multi-channel terrain information fusion, and multi-scale terrain information fusion for saddle decision.

[0124] Step 3.1: The terrain information parallel extraction stage includes feature extraction and parallel calculation. Different levels of neural networks can extract different feature information. Shallow neural networks can extract texture, boundary and other low-level visual information, and deep neural networks can extract deep-level semantic information. In the application of multi-feature fusion of neural networks, deep semantic features are usually used for fusion. Therefore, in the parallel feature extraction step of each channel, multi-level neural networks are used to extract deep-level semantic feature fusion after the original data is extracted. As shown in Figure 6(a) As shown, three convolution kernels are used to extract the input features of each channel, and a small convolution kernel of 3*3 is used (Table 1), because small convolution kernels are more conducive to local feature preservation. Convolution kernel 1 and convolution kernel 2 are used to extract shallow visual features, and convolution kernel 3 is used to extract deep semantic features. Finally, resampling is performed to keep the scale of the fused features consistent, and the feature maps are resampled to 4*4 in size. Before resampling, the input feature sizes of each channel are inconsistent, and the feature levels do not meet the requirements for feature fusion. Therefore, as shown in Figure 6 (a) Before resampling, five independent parallel convolutional neural networks are set up for each scale, and a total of 20 parallel convolutional neural networks are used for feature extraction. This process can be expressed as:

[0125]

[0126] wherein represents the i-th channel topographic information of scale s after convolution calculation of multi-channel topographic information, represents the i-th topographic feature of scale s.

[0127] Table 1

[0128]

[0129] Table 2

[0130]

[0131] Step 3.2: Multi-channel topographic information fusion. As shown in Figure 6 (b) and Table 2, first, the deep topographic semantic information of each channel after resampling is combined, and then two layers of convolution operations are performed, with a small convolution kernel of 3*3. This can be expressed as:

[0132]

[0133] Step 3.3: Saddle decision of multi-scale topographic information fusion. As shown in Figure 6 (c) This step includes three hidden layers. The topographic information of each scale is spliced through the channel, then flattened, and then input into three fully connected layers to output the judgment result of the saddle point

[0134] y = Softmax (Fully_connect (Concatenate (FC4 00 , FC 800 , FC 1200 , FC 1600 ))

[0135] Step 3.4: Activation function design. To improve the network's nonlinear modeling capabilities, activation functions need to be added to each layer. Neural networks can only learn hierarchical nonlinear mapping with activation functions. Therefore, the Leaky Relu function is chosen as the activation function of each layer to ensure that all negative values have a non-zero slope, avoiding the premature appearance of gradient vanishing or gradient explosion during training. The specific calculation method is as follows:

[0136]

[0137] where y i is the output value after the activation function, x i is the output value after the weight multiplication of each layer, and a is a constant between [0, 1), usually set to 0.5.

[0138] Step 3.5: Loss function design. The loss function is used in the model training phase to quantify the difference between the training results and the actual label results, and to adjust the weights in the model based on the loss. Here, the model output is the probability of keeping or not keeping the candidate saddle, which can be considered as a binary classification problem. Therefore, the binary cross-entropy is chosen as the loss function, and its calculation method is as follows:

[0139]

[0140] where Loss is the total loss function; N is the number of samples; K is the number of categories; x ij is the output value of the fully connected layer, f(x ij ) is the normalized probability after the softmax activation function; and y ij is the label probability.

[0141] Thus, by taking the target saddle point as the center, extracting a plurality of terrain parameters under a preset area can comprehensively capture the terrain features of the saddle point; these features are extracted at multiple scales, which helps to cover the control area and terrain complexity of different saddle points, and provides rich input data for subsequent neural network training; the extracted terrain features are used to train the convolutional neural network, so that the network can learn the multi-scale terrain features of the saddle point; through the preset activation function and loss function, the weights and biases of the network are optimized, and the recognition ability of the network for the saddle point is improved; the multi-scale terrain feature fusion mechanism enables the network to comprehensively consider terrain information at different scales, improving the accuracy and robustness of recognition; the trained convolutional neural network can distinguish candidate saddle points to distinguish true and false saddle points; through the output result of the network, false saddle points in the candidate saddle point set can be eliminated, and the final saddle point set is obtained. According to the saddle point extraction method proposed in the embodiments of the present application, DEM data used for terrain saddle extraction is obtained; based on the preset positive and negative terrain method, a candidate saddle point set is extracted from the DEM data, and a plurality of terrain features of each candidate saddle point in the candidate saddle point set at multiple scales are extracted; the plurality of terrain features of each candidate saddle point at multiple scales are input into a pre-established multi-scale terrain feature fusion convolutional neural network, and false saddle points in the candidate saddle point set are eliminated according to the output result to obtain the final saddle point. Thus, by combining the positive and negative terrain method and the multi-scale feature fusion convolutional neural network, the problems of excessive false saddle points when extracting saddle points by the positive and negative terrain method are solved, the extraction precision and accuracy of the saddle point are improved, and the intelligent analysis and recognition of the terrain and topography are promoted.

[0142] Next, the saddle point extraction device according to the embodiments of the present application is described with reference to the accompanying drawings.

[0143] Figure 7 is a block schematic diagram of the saddle point extraction device according to the embodiments of the present application.

[0144] As shown in Figure 7 , the saddle point extraction device 10 includes an acquisition module 100, an extraction module 200, and a determination module 300.

[0145] The acquisition module 100 is configured to acquire DEM data used for terrain saddle extraction.

[0146] The extraction module 200 is configured to extract a candidate saddle point set from the DEM data based on a preset positive and negative terrain method, and extract a plurality of terrain features of each candidate saddle point in the candidate saddle point set at multiple scales.

[0147] The determining module 300 is configured to input the multiple terrain features of each candidate saddle point at multiple scales into a pre-established convolutional neural network for fusing multiple-scale terrain features, and remove false saddle points in the candidate saddle point set according to an output result to obtain final saddle points.

[0148] Optionally, the extracting module 200 is specifically configured to: determine inverse DEM data according to the DEM data, extract a ridge line based on the DEM data, and extract a mountain control line based on the inverse DEM data; determine intersection data of the ridge line and the mountain control line, and pre-process the intersection data to obtain a candidate saddle point set.

[0149] Optionally, the determining module 300 is further configured to: obtain a target saddle point set for training; extract multiple terrain parameters in a preset area with each saddle point in the target saddle point set as a center to obtain terrain features of each saddle point in the target saddle point set; determine saddle control areas at multiple scales, and train the preset convolutional neural network by using the terrain features of each saddle point in the target saddle point set based on a preset activation function, a preset loss function and the saddle control areas at multiple scales to obtain the pre-established convolutional neural network for fusing multiple-scale terrain features.

[0150] Optionally, the preset activation function is:

[0151]

[0152] wherein y i is an output value after the activation function, x i is an output value after weight multiplication at each level, and a is a constant between 0 and 1.

[0153] Optionally, the preset loss function is:

[0154]

[0155] wherein Loss is a total loss function, N is a sample quantity, K is a classification quantity, x ij is an output value of a full connection layer, i is an i-th sample, j is a j-th classification result (including two classification results of saddle and false saddle), f(x ij ) is a probability after normalization of a softmax activation function, and y ij is a label probability.

[0156] It should be noted that the foregoing explanation and description of the saddle point extraction method embodiment are also applicable to the saddle point extraction device of the embodiment, which will not be described herein again.

[0157] According to the saddle point extraction device provided by the embodiment of the present application, DEM data used for terrain saddle extraction is obtained; based on a preset positive and negative terrain method, a candidate saddle point set is extracted from the DEM data, and a plurality of terrain features of each candidate saddle point in the candidate saddle point set at a plurality of scales are extracted; the plurality of terrain features of each candidate saddle point at the plurality of scales are input into a convolutional neural network for multi-scale terrain feature fusion which is established in advance, and a false saddle point in the candidate saddle point set is eliminated according to an output result, so as to obtain a final saddle point. In this way, by combining the positive and negative terrain method and the multi-scale feature fusion convolutional neural network, the problem of excessive false saddle points when the positive and negative terrain method extracts the saddle points is solved, the extraction precision and accuracy of the saddle points are improved, and intelligent analysis and identification of the terrain and landform are promoted.

[0158] Figure 8 The structure schematic diagram of the electronic device provided by the embodiment of the present application is provided. The electronic device can include:

[0159] The memory 801, the processor 802 and the computer program stored in the memory 801 and executable on the processor 802.

[0160] The processor 802 implements the saddle point extraction method provided in the above embodiment when executing the program.

[0161] Further, the electronic device further includes:

[0162] The communication interface 803 is used for communication between the memory 801 and the processor 802.

[0163] The memory 801 is used to store the computer program executable on the processor 802.

[0164] The memory 801 can include a high-speed RAM memory, and can also include a non-volatile memory, for example, at least one disk memory.

[0165] If the memory 801, the processor 802 and the communication interface 803 are independently implemented, the communication interface 803, the memory 801 and the processor 802 can be connected to each other through a bus and complete communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 8Only one bus or only one type of bus can exist, however.

[0166] Optionally, in a specific implementation, if the memory 801, the processor 802 and the communication interface 803 are integrated on a chip, the memory 801, the processor 802 and the communication interface 803 can complete the communication among each other through an internal interface.

[0167] The processor 802 can be a central processing unit (CPU) or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement one or more embodiments of the application.

[0168] The embodiment of the application further provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the saddle point extraction method.

[0169] The embodiment of the application further provides a computer program product, which stores a computer program, and the program is executed by a processor to implement the saddle point extraction method.

[0170] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or N embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples without contradiction.

[0171] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the application, the meaning of "N" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0172] Any process or method described in a flowchart or otherwise described herein can be understood as representing a module, segment, or portion of code that includes one or more executable instructions for implementing the specified logical functions or steps, and the preferred embodiments of the application include additional or fewer steps, or combinations of steps, or the order of the steps can be different from those shown or discussed. It is intended that additional or fewer steps be performed between any two steps, and one of ordinary skill in the art would recognize many variations based on the functional description.

[0173] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the above embodiments, the steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. As such, if desired, the functions can be implemented in hardware, using, for example, any or a combination of the following technologies, which are all well-known in the art: a discrete logic circuit having logic gates for implementing logic functions upon an application of data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0174] Those of skill in the art would understand that information and signals can be represented using any of a variety of technologies and techniques. For example, data, instructions, commands, information, signals, bits, symbols, and chips that can be referenced throughout the above description can be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof. Those of skill in the art would further appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality, and can be implemented with any number of hardware or software components configured to perform the stated functionality. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality, and can be implemented with any number of hardware or software components configured to perform the stated functionality.

Claims

1. A saddle point extraction method characterized by, The method comprises the following steps: acquiring DEM data for terrain saddle extraction; extracting a candidate saddle point set from the DEM data based on a preset positive and negative terrain method, and extracting a plurality of terrain features of each candidate saddle point in the candidate saddle point set at a plurality of scales; inputting the plurality of terrain features of each candidate saddle point at a plurality of scales into a pre-established convolutional neural network for multi-scale terrain feature fusion, and removing false saddle points in the candidate saddle point set according to an output result to obtain a final saddle point; wherein the pre-established convolutional neural network for multi-scale terrain feature fusion comprises terrain information parallel extraction, multi-channel terrain information fusion and multi-scale terrain information fusion saddle decision; the terrain information parallel extraction comprises feature extraction and parallel calculation, and different feature information is extracted by a neural network at different levels, wherein shallow neural networks extract low-level visual information such as texture and boundary, and deep neural networks extract deep-level semantic information; in multi-feature fusion application of the convolutional neural network, deep semantic features are used for fusion; before resampling, five independent parallel convolutional neural networks are separately set for each scale, and a total of 20 parallel convolutional neural networks are used for feature extraction; after the deep terrain semantic information of each channel with consistent size after resampling is combined, two convolution operations are performed, wherein a 3*3 small convolution kernel is used as the convolution kernel; the multi-scale terrain information fusion saddle decision comprises three hidden layers; the terrain information at each scale is spliced through a channel, flattened and then input into three fully connected layers to output a judgment result of the saddle point; a Leaky Relu function is used as an activation function; a binary cross entropy is used as a loss function, and each weight value in the model is adjusted.

2. The method of claim 1, wherein, The candidate saddle point set is extracted from the DEM data based on the preset positive and negative terrain method, which comprises: determining inverse DEM data according to the DEM data, extracting a ridge line based on the DEM data, and extracting a mountain control line based on the inverse DEM data; determining intersection data of the ridge line and the mountain control line, and pre-processing the intersection data to obtain the candidate saddle point set.

3. The method of claim 1, wherein, Before the plurality of terrain features of each candidate saddle point at a plurality of scales are input into the pre-established convolutional neural network for multi-scale terrain feature fusion, the method further comprises: acquiring a target saddle point set for training; extracting a plurality of terrain parameters in a preset area with each saddle point in the target saddle point set as a center to obtain terrain features of each saddle point in the target saddle point set; determining a plurality of scale saddle control areas, and training the pre-established convolutional neural network by using the terrain features of each saddle point in the target saddle point set based on a preset activation function, a preset loss function and the plurality of scale saddle control areas to obtain the pre-established convolutional neural network for multi-scale terrain feature fusion; wherein the scale saddle control area is set based on an area of a same horizontal plane of a saddle and two mountains at a saddle elevation.

4. The method of claim 3, wherein, The preset activation function is: ； wherein, is the output value after the activation function, is the output value after the weight multiplication of each level, is a constant between [0, 1).

5. The method of claim 3, wherein, The preset loss function is: ； wherein Loss is a total loss function, N is a number of samples, and K is a number of classifications, is a value output by a fully connected layer, is a value output by a fully connected layer, i is a value output by a fully connected layer, is a value output by a fully connected layer, j is a value output by a fully connected layer, is a value output by a fully connected layer, is a value output by a fully connected layer.

6. A saddle point extraction apparatus characterized by comprising: comprises: An acquisition module is configured to acquire DEM data for terrain saddle extraction; An extraction module is configured to extract a candidate saddle point set from the DEM data based on a preset positive-negative terrain method, and extract a plurality of terrain features of each candidate saddle point in the candidate saddle point set at a plurality of scales; A determination module is configured to input the plurality of terrain features of the each candidate saddle point at the plurality of scales into a pre-established convolutional neural network for multi-scale terrain feature fusion, and remove false saddle points in the candidate saddle point set according to an output result to obtain a final saddle point; wherein the pre-established convolutional neural network for multi-scale terrain feature fusion includes terrain information parallel extraction, multi-channel terrain information fusion, and saddle decision based on multi-scale terrain information fusion; the terrain information parallel extraction includes feature extraction and parallel calculation, and different feature information is extracted by neural networks at different levels, wherein shallow neural networks extract low-level visual information such as texture and boundary, and deep neural networks extract deep-level semantic information; in multi-feature fusion application of the convolutional neural network, deep semantic features are used for fusion; before resampling, five independent parallel convolutional neural networks are separately set for each scale, and a total of 20 parallel convolutional neural networks are used for feature extraction; after the deep terrain semantic information of each channel with the same size after resampling is combined, two convolution operations are performed, wherein a 3*3 small convolution kernel is used as the convolution kernel; the saddle decision based on multi-scale terrain information fusion includes three hidden layers; terrain information at each scale is spliced through a channel, flattened, and then input into three fully connected layers to output a judgment result of a saddle point; a Leaky Relu function is used as an activation function; a binary cross-entropy is used as a loss function, and each weight value in the model is adjusted.

7. The apparatus of claim 6, wherein, The extraction module is specifically configured to: determine inverse DEM data according to the DEM data, extract a ridge line based on the DEM data, and extract a mountain control line based on the inverse DEM data; determine intersection data of the ridge line and the mountain control line, and pre-process the intersection data to obtain the candidate saddle point set.

8. An electronic device, comprising: comprise: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the saddle point extraction method according to any one of claims 1-5.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the saddle point extraction method according to any one of claims 1-5.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the saddle point extraction method according to any one of claims 1-5.

Images