CN114359701A - Mountain vertex recognition algorithm based on deep learning - Google Patents

Abstract

The invention discloses a mountain vertex algorithm based on deep learning, and belongs to the field of mechanical learning. According to the method, a ResNet-101 network is constructed under Keras, a residual network is constructed to screen mountain samples, mountain vertexes are screened and positioned through an RPN module and a regression network, and are finally mapped to DEM data to determine specific coordinates of the mountain vertexes, and finally mountain vertex types can be further calculated.

Description

Mountain vertex recognition algorithm based on deep learning

Technical Field

The invention belongs to the field of mechanical learning, and particularly relates to a mountain vertex recognition algorithm based on deep learning.

Background

The method mainly comprises four categories of geometric analysis, section elevation extreme value, hydrological basin analysis and topological structure analysis, wherein the prior commonly used method for extracting the mountain top points is a method for extracting the mountain top points by researching an actor to put forward an inverse terrain DEM, thereby improving the automatic extraction precision and efficiency of the mountain top points to a certain extent, and also provides a maximum relief threshold index by researching the actor to provide reference for realizing the accurate mathematical description of the mountain top points, then uses a point-grid correlation method to explore the grid characteristics among the terrain characteristic points to obtain conclusions that the spatial distribution characteristics of the mountain top points have correlation with the terrain types, and the like, and finally triggers from the geometric characteristics of the terrain characteristic points to research the algorithm for extracting the mountain top points through the DEM, but the algorithm inevitably has terrain description errors and terrain relief changes, so that a large number of pseudo-mountain vertexes exist in the extracted result of the algorithm.

The research on the mountain top points plays an important role in various geological exploration researches and remote sensing terrain researches, and can provide data support for tourism industry development and various sports projects which need to depend on mountain bodies, and has very important significance on the construction, maintenance and safety guarantee of ski fields through accurate identification of the mountain top points, so that the research on the mountain top points is not practically useless.

The accuracy and efficiency of identifying and extracting the mountain vertexes by using the traditional mountain vertex extraction methods are not high, the number of extracted pseudo mountain vertexes is large, and the error of the obtained data is large.

Disclosure of Invention

Aiming at the problems in the prior art, the invention aims to provide a mountain vertex identification algorithm based on deep learning, and the problems of low mountain vertex identification accuracy and low efficiency in the prior art are solved.

In order to solve the problems, the technical scheme adopted by the invention is as follows:

a mountain vertex recognition algorithm based on deep learning can accurately extract mountain vertices aiming at complex terrains, and the method comprises the following steps:

step 1: dividing the landform type of a target area, obtaining the relief degree and elevation value data, determining the lowest elevation threshold value of a mountain vertex and establishing a learning sample set;

step 2: constructing a ResNet-101 network under Keras to construct a residual network, and constructing a mountain top point identification frame through the residual network to identify mountain top area morphological characteristics to primarily screen out mountain samples;

and step 3: searching a target mountain top area possibly existing in the mountain land sample through an RPN network module;

and 4, step 4: further screening the mountain top area extracted by the RPN module through a classification regression network, and mapping the mountain top area to the original DEM data;

and 5: and further distinguishing the mountain top types by using DEM data obtained after mapping.

The undulation calculation formula is as follows:

where m and n represent the number of grids, d represents the degree of relief of the region,

average elevation value, h, of representative area_xyThe elevation values of the grid points.

The ResNet-101 network model carries out preliminary training through ImageNet data, fine-tunes parameters through a self-built mountain top sample book, and finally carries out mountain top feature extraction in a transfer learning mode.

The RPN network module adopts a mean shift clustering algorithm to mark samples according to the mountain tops, sets anchor frame parameters suitable for a mountain top target area, then completes positioning of the center of each group/class by updating the candidate point of the center point to the mean value of the points in the sliding window based on the centroid, further screens the candidate windows to form corresponding groups of the center point set, then performs clustering analysis on the samples, and determines the anchor frame finally used for mountain top identification.

The regression network inputs final classification layer and regression layer calculation after performing maximum pooling on the target mountain top area samples obtained by the RPN module, outputs probability value that each candidate sample meets mountain top characteristics, then obtains elevation value of the sample, and then obtains the elevation value of the sample through a formula

Calculating the exact coordinates of the mountain vertex, where M_ijIs the position of the mountain vertex, a_ijIs the elevation value at coordinate (i, j).

The method for judging the mountain top type comprises the following steps of;

1) determining the highest communication area where the mountain top point is contracted and calculating the area S;

2) judging whether S is smaller than a given minimum threshold value or not, and distinguishing sharp mountain tops;

3) calculating the undulation degree of all sample points except the hill top obtained in the last step;

4) judging whether the undulation degree is less than a given minimum threshold value, and distinguishing flat mountain tops from round mountain tops

Compared with the prior art, the method has the advantages that the algorithm is pre-trained and migrated and learned through the deep learning network to screen and identify the mountain top points, accuracy of algorithm identification can be remarkably improved, errors can be reduced, identification of false mountain top points can be reduced, identified mountain top types can be accurately distinguished through the algorithm, and influence of manual selection and processing software on mountain top element areas can be effectively avoided.

Drawings

FIG. 1 is a flow chart of an algorithm;

FIG. 2 is a DEM data processing flow chart

FIG. 3 is a rendering of contour lines in a learning sample set PEAK-100;

FIG. 4 is a chongy district elevation distribution diagram;

FIG. 5 is a schematic diagram of the undulation degree of the Chongy district;

FIG. 6 is a Chongy district terrain distribution degree;

FIG. 7 is a schematic diagram of a residual module structure;

FIG. 8 is a distribution diagram of mountain top points in a Chongli area;

FIG. 9 is an enlarged view of a portion of FIG. 8;

FIG. 10 is a map of local mountain peaks;

FIG. 11 shows the mountain top extraction result of the algorithm;

FIG. 12 shows the mountain vertex type extraction result of the algorithm;

FIG. 13 is a comparison chart of mountain top extraction for whether landform classification is used;

FIG. 14 is a contour plot mountain apex recognition plot;

FIG. 15 is a mountain top recognition chart of a gray scale;

fig. 16 is a mountain peak recognition diagram obtained by superimposing the gray-scale maps of the contour maps.

Detailed Description

The invention is further described with reference to specific examples.

Example 1

The Chongli area of Zhangkou city in Hebei province is located in the northwest part of Hebei province, is located in the transition zone between inner Mongolian plateau and North China plain, and belongs to a transition type mountain area under a dam on a dam, the mountains are more and seven ski resorts are built, so the topography of the Chongli area is adopted as the implementation object of the embodiment.

In the field of geographic information, contour topographic maps mainly reflect the landform forms of mountain elements, and the spatial elevation changes of the topographic elements are difficult to accurately describe in an image form. The general flow of the invention is shown in fig. 1, and the digital image reflects the elevation change of the mountainous region with multi-level gray difference by performing gray rendering processing on the DEM data. Therefore, in order to effectively utilize the depth network to extract the topographic features in the DEM data, the topographic features of the mountain land data are reflected by the contour line image, the elevation change trend in the mountain land data is represented by the gray level rendering graph, as shown in fig. 2, contour line generation and gray level rendering processing are performed on the original DEM data, the mountain top features under the two modes are fused in the form of image synthesis, and then learning sample data which can be received by the depth network are formed.

Before deep learning, a sample set for deep learning needs to be built, in this embodiment, the DEM data provided by ASTER GDEM is used as basic data, mountain land parts are selected according to elevation distribution, 1000 mountain land data blocks with the size of 320 × 320 are cut out, it is ensured that each data block has at least one or more mountain tops, and the DEM data is preprocessed according to the method shown in fig. 2. Simultaneously, a method of minimum external rectangles is adopted to mark mountain top areas in the superposed images, each marked rectangular frame is guaranteed to have only one mountain top, background information is contained as little as possible, mountain top sample data is formed, and a superposed mountain contour rendering graph is shown in FIG. 3;

X1	X2	X3	X4	label
					370	120	430	200	1
325	240	370	370	1
					300	420	250	440	1

the sample label information in the superscript, then in turn, is the left lower corner coordinates (x1, y1) (x1, y1) and the right upper corner coordinates (x2, y2) (x2, y2) of the mountain top region coordinate frame and the cable label. Finally, a learnable sample set containing a mountain top target area is established and named as PEAK-100, 3500 mountain top samples are included, and the mountain top sample set is divided into a training set, a verification set and a test set according to the ratio of 6: 3: 1.

The waviness (Undulation) is an extremely important value in the algorithm, and refers to the relative change magnitude of the elevation of the area, namely the standard deviation in mathematical definition. Therefore, for raster data of size m × n, the undulation degree of the region is calculated by:

h_xyrepresenting the elevation values of the grid points, d representing the relief of the area,

the average elevation value of the representative area is obtained through the formula, and then the table is divided through the basic landform form:

and gradually dividing the basic landform of the Chongli area according to two indexes of the surface relief degree and the altitude (altitude).

The elevation distribution diagram of the Chongli area is shown in FIG. 4, the relief map is shown in FIG. 5, the landform division of the Chongli area is shown in FIG. 6 by combining the above formula and the table, the obtained FIG. 5 is compared with the result of the actual research, and the division result is consistent with the actual situation of the Chongli area.

The following table is obtained by combining the formula, the landform type division marking and the setting principle of the minimum elevation of the mountain top point:

determining the lowest elevation threshold value P of the peak of the Chongli area according to the table_min

In order to extract coordinates of mountain vertices more accurately, mountain vertex features need to be extracted, in order to improve the recognition effect of mountain vertex regions, a ResNet-101 network is constructed under Keras, a residual network is constructed to solve the problem of gradient degradation in a deep learning network, an unknown function h (x) identity which needs to be learned and approximated is mapped into a function which approximates f (x) h (x) -x, and the structure of residual modules is shown in fig. 7, wherein two Convolution Layers (CL) represent that a secondary Convolution operation is performed, x is an input value of a residual block, f (x) is an input value of a Convolution block, and f (x) + x is an output of the residual block.

Considering the limitation of data of mountain top sample samples in a self-built Chongli area, the performance of a convolutional neural network is influenced by the sample size which can be learned, therefore, an ImageNet data set is adopted to carry out preliminary training on a ResNet-101 network model, bottom layer volume blocks which are extracted and related to thousands of layers of general features are frozen, then the self-built mountain top sample samples are used for carrying out internal parameter fine tuning on the top layer volume blocks, and the feature extraction of mountain tops is carried out in a migration learning mode.

After the extraction of the features of the mountain top is finished, a target mountain top area which is possibly existed in the mountain sample is searched through a kernel RPN network module of the Faster R-CNN, a common RPN module performs space convolution on the feature map in a sliding block mode, deeper features of the mountain top area are selected, and meanwhile, each pixel point on the obtained space convolution map is mapped to the position of the original mountain sample to generate anchor frames with different size ratios. In order to improve the identification accuracy of mountain top pixels, in the embodiment, a mean shift clustering algorithm is adopted to set anchor frame parameters suitable for a mountain top target area according to mountain top point marking samples, wherein the main idea of the mean shift clustering algorithm is to find out a dense area of data points based on an algorithm of a sliding window, then, based on a centroid, the candidate points of a central point are updated to be the mean value of points in the sliding window to complete the positioning of the central point of each group/class, further, similar windows of the candidate windows are removed, a central point set and corresponding groups are finally formed, the size of a marking frame in a self-built mountain top training set is used as the input of the clustering algorithm, and the anchor frame finally used for mountain top point identification is determined by performing cluster analysis on the marking frame in the sample set.

Next, the anchor frame determined by the RPN module is screened again, by calculating an Intersection-over-Union ratio (IoU) between the anchor frame and a mountain top region mark frame, and setting IoU >0.6 as a foreground threshold, IoU <0.4 as a background threshold, dividing the anchor frame into a foreground threshold containing a mountain top and a background threshold not containing the mountain top, discarding the anchor frame at the middle threshold, screening candidate frames by using a non-maximum suppression NMS method, sorting the remaining anchor frames according to a ranking order, and retaining the candidate frames with a higher probability of the mountain top to realize rough estimation positioning of the mountain top in the Chongy region, wherein the calculation mode of IoU is shown as the following formula:

wherein anchor represents an anchor frame and GT is a mark frame of the mountain top area.

Performing regression calculation on the position of the anchor frame by using the formula (2), and obtaining the offset [ d ] of the predicted mountain top candidate region after the RPN network training convergence_x(A),d_y(A),d_w(A),d_h(A)]

In the above formula [ B_x,B_y,B_w,B_h]Denotes the center coordinates and length and width of the anchor frame, [ G ]_x,G_y,G_w,G_h]The center coordinates and the length and width of the real mountain top mark frame are represented, and the approximate position of the prediction candidate mountain top area can be calculated according to the offset.

After the candidate regions are screened, the candidate regions need to be identified and corrected through a classification regression network, and due to the fact that the sizes of anchor frames of the candidate regions are different, calculation of an input final classification layer and a regression layer needs to be conducted after maximum pooling is conducted, and probability values of mountain top features, which are met by each candidate frame, are output.

Finally, a target anchor frame and mountain tops in the target anchor frame are obtained, then the mountain tops are required to be mapped into original DEM data, the elevation values of all coordinates in a mountain top area are obtained through an Arcgis10.8 platform, accurate mountain top coordinates can be marked by calculating the maximum elevation value points in the area, and the calculation method is as follows:

wherein M is_ijIs the position of the mountain vertex, a_ijIs the elevation value at coordinate (i, j).

After the information of the mountain tops of the area is obtained, the types of the mountain tops of the area need to be further identified, and according to the analysis of the mountain tops, the mountain tops can be roughly divided into three types, namely a sharp mountain top, a round mountain top and a flat mountain top, wherein the sharp mountain top is sharp in shape, gradually becomes point-shaped, the area of the mountain top is small, the area of the part of the mountain top of the round mountain top is large, the whole shape is relatively smooth, and the flat mountain top is flat in top, so that the position of the mountain top is difficult to determine. Firstly, determining the highest communication area where the mountain top points are located and calculating the area S of the highest communication area, judging whether S is smaller than a given minimum threshold value, if so, judging the top of the sharp mountain, further judging whether the size of the remaining target mountain top points in the last step is smaller than the given minimum threshold value, if so, judging the top of the flat mountain, and if so, judging the top of the round mountain.

At this point, all the processes of the algorithm are finished, and the following is the actual calculation effect in the embodiment.

In the embodiment, the Keras deep learning development environment is adopted, the self-built mountain top sample data set PEAK-100 is used for network training, the initial learning rate of the model is 0.002, the momentum is 0.8, and the iteration times are 10000. And setting the initial size parameter of the anchor frame by using a K-means clustering algorithm to mark frames in the mountain top training sample, respectively calculating the width and the height of the mark frame through the coordinates of the upper left corner and the lower right corner of the mountain top area mark frame, using the mark frame as a data sample of the clustering algorithm, and setting a criterion function of the K-means clustering algorithm as follows:

in the formula J_cIs a common clustering criterion function, S_jFor class j sample summation, m_jIs S_jThe sample mean of (1).

Setting j to 0.02, performing cluster analysis on the training samples by using c to 4, and setting a coordinate parameter corresponding to a cluster center as an initial anchor frame size of the fast R-CNN network according to a result obtained by the cluster analysis, namely replacing an original anchor frame parameter by four length-width ratios of (42.85, 70.65), (65.45, 83.67), (92.54, 123.52) and (132.45, 194.85) to be used as an initial anchor frame setting value of each mountain top area target.

2792 mountain tops obtained according to the experiment are shown in FIG. 8, and as can be seen from FIG. 9 and the enlarged partial view 10, the obtained mountain tops have high conformity with the topography of the Chongli area, and further research on the topography of the Chongli area has high reference value. Compared with the experiment that 1680 mountain top points are obtained by a traditional manual identification method, the accuracy rate of the experiment is improved by 67%, and the classification statistics of the mountain top points in the Chongli area is shown in FIG. 11 and FIG. 12 according to the landform type and the mountain top point type of the Chongli area.

Example 2

Based on the embodiment 1, we directly extract the mountain vertices without performing the terrain classification to obtain the local map of the mountain vertices extracted without using the terrain classification and the comparison map of the mountain vertices extracted with using the terrain classification, as shown in fig. 13 (left is unused and right is used), it can be seen that a significant increase occurs when no terrain classification is used, and the reason is that the lowest elevation of the mountain vertices in the local area is accurately determined without using the terrain classification.

In order to further evaluate the mountain vertex extraction result, the false extraction rate and the missed extraction rate of mountain vertex identification are used as model evaluation indexes. Wherein the false extraction rate R_iRepresenting the ratio of the number of false positives in the identified mountain vertices; leakage extraction rate R_oIndicating the ratio of the number of false positives in the identified mountain vertices. The formula of the calculation formula is as follows:

in the formula, CP is the correct number of mountain tops in the recognition result; the IP is the number of mountain top points extracted by mistake in the identification result; and the number of the top points of the missed lifting mountain in the OP identification result.

In the preliminary data processing stage, the used DEM mountain vertex data is obtained by fusing the contour line of the Chongli area with the gray level map. In order to verify the rationality and effectiveness of the method, the identification effects of the mountain tops under different conditions are compared, and the experimental results are shown in fig. 14 to 16, wherein fig. 14 is the identification result of the mountain tops in a contour diagram, fig. 15 is the identification result of the mountain tops in a gray scale diagram, and fig. 16 is the identification result of the mountain tops after the contour diagram and the gray scale diagram are superimposed, wherein a black frame and a red frame respectively represent missed extraction and false extraction. The test experiment is carried out through the self-built mountain vertex sample set, and the statistical results are shown in the following table:

	contour map	Grey scale map	Overlay map
				Ri(％)	15.41	7.52	4.25
Ro(％)	12.05	9.23	4.15

In the method, the mountain top point information acquired according to the contour map has a missed-lifting phenomenon compared with the mountain top point result acquired after superposition processing, the mountain top point information acquired according to the gray scale map has a false-lifting phenomenon compared with the mountain top point result acquired after superposition processing, the contour map mainly reflects the landform form of mountain data, the gray scale map mainly reflects the elevation change trend in the mountain data, and the extraction of the mountain top form by considering the gray scale map and the contour map in one way can be seen by combining the above table experiment results, so that a candidate frame is difficult to accurately mark a mountain top target area, the missed-lifting and the multiple-lifting phenomena can occur, and the goal of improving the mountain top point extraction efficiency can be achieved by fusing the gray scale map and the contour map of the Chongli area.

According to the method, firstly, a residual error network is utilized to carry out deep excavation on morphological characteristics of a mountain top area in a mountain land sample, then a core RPN network module of fast R-CNN is combined to automatically generate a high-quality mountain top target area, finally the mountain top target area in a sample image is mapped back to original DEM data, and the position of an elevation maximum value in the target area is calculated to serve as a final mountain top point coordinate.

According to the mountain vertex recognition algorithm disclosed by the invention, from the test result, the false extraction rate of the mountain vertex is reduced to 4.25%, the extraction omission rate is reduced to 4.15%, the recognition accuracy rate is improved to 92%, compared with the traditional method, the accuracy is improved by 67%, the problem of high cost and large error of manual recognition is solved, and more perfect data are provided for various researches and industries needing mountain vertex values.

Claims (6)

1. A mountain vertex recognition algorithm based on deep learning is characterized by comprising the following steps:

step 1: dividing the landform type of a target area, obtaining the relief degree and elevation value data, determining the lowest elevation threshold value of a mountain vertex and establishing a learning sample set;

step 2: constructing a ResNet-101 network under Keras, constructing a residual network, and constructing a mountain top area morphological characteristic identification framework through the residual network to preliminarily screen mountain land samples;

and step 3: searching a target mountain top area possibly existing in the mountain land sample through an RPN network module;

and 4, step 4: further screening the mountain top area extracted by the RPN module through a classification regression network, and mapping the mountain top area to the original DEM data;

and 5: and further distinguishing the mountain top types by using DEM data obtained after mapping.

2. The mountain vertex recognition algorithm based on deep learning of claim 1, wherein the undulation degree is calculated by the following formula:

where m and n represent the number of grids, d represents the degree of relief of the region,

average elevation value, h, of representative area_xyElevation values of grid points

3. The mountain vertex recognition algorithm based on deep learning of claim 1, wherein the ResNet-101 network model is pre-trained through ImageNet data, parameters are fine-tuned through a self-built mountain vertex sample book, and finally, a mountain vertex feature extraction is performed in a migration learning mode.

4. The mountain vertex recognition algorithm based on deep learning of claim 1, wherein the RPN network module sets anchor frame parameters suitable for a mountain vertex target area according to mountain vertex labeling samples by using a mean shift clustering algorithm, then completes positioning of the center of each group/class by updating candidate points of the center point to the mean of points in a sliding window based on the centroid, further screens the candidate windows to form corresponding groups of the center point set, then performs cluster analysis on the samples, and determines the anchor frame finally used for mountain vertex recognition.

5. The mountain vertex recognition algorithm based on deep learning of claim 1, wherein the regression network inputs a final classification layer and a regression layer calculation after performing maximum pooling on the target mountain vertex area samples obtained by the RPN module, outputs a probability value that each candidate sample meets mountain vertex features, then obtains an elevation value of the sample, and then obtains the elevation value of the sample through a formula

Calculating the exact coordinates of the mountain vertex, where M_ijIs the position of the mountain vertex, a_ijIs the elevation value at coordinate (i, j).

6. The mountain top point identification algorithm based on deep learning of claim 1, wherein the method for determining the mountain top type is;

1) determining the highest communication area where the mountain top point is contracted and calculating the area S;

2) judging whether S is smaller than a given minimum threshold value or not, and distinguishing sharp mountain tops;

3) calculating the undulation degree of all sample points except the hill top obtained in the last step;

4) and judging whether the undulation degree is smaller than a given minimum threshold value, and distinguishing flat mountain tops from round mountain tops.

Images