CN111768039B - Animal home domain estimation method based on active learning - Google Patents

Abstract

The invention discloses an animal home domain estimation method based on active learning, which takes a GPS point set, basic geographic information data and a plurality of geographic information data of DEM data as input data, wherein the GPS point position data set is tracking data of animal activities, obstacles can be determined by the basic geographic information data, and the obstacles of different animals are possibly different; digital Elevation Model (DEM) data is used for determining the influence of terrain on the animal activity, and then a cost distance calculation module is used for calculating the cost distance from each GPS tracking point to any point in the activity area of the animal; calculating a fuzzy probability distribution of the animal moving in the area by using a probability distribution calculation module; the home domain extraction module is used for extracting a core area and a non-core area (boundary area) of the animal activity range by determining a threshold value in advance, so that the calculation of the home domain is realized.

Description

Animal home domain estimation method based on active learning

The application obtains the natural science fund (17JCZDJC 39790) in Tianjin and the national natural science fund (41971410) subsidizes.

Technical Field

The invention relates to a fuzzy geographic phenomenon modeling method in geographic information science and an animal home domain determination method in ecology and animal behavior science, in particular to a method for acquiring the position of an animal by using a GPS tracker and estimating the moving range of the animal by using position data, basic geographic information data, DEM data and the like.

Background

The estimation of the home domain (range of motion of animals) is a central topic of spatial ecology research and is also the basis for understanding animal behaviors. The home domain of an animal is generally defined as the percentage of distributed area coverage that encompasses all possible locations. Many complex factors affect the location and extent of an animal's home domain, such as the animal's behavioral characteristics, geographic environment, body type, and population range; therefore, estimation of the home domain remains a challenge in the field of ecology. The performance of the home domain estimation method can be currently evaluated by the following criteria: (1) the probability or likelihood of animal distribution can be naturally expressed; (2) the range of the home area can avoid obstacles such as cliffs, mountains, rivers and roads for certain types of animals; (3) the influence of the terrain can be considered; (4) the estimated home domain range error is minimal.

Common estimation methods include:

(1) the minimum convex polygon method, which is one of the earliest and simplest estimation methods, calculates the range of a home domain by drawing a convex polygon around a position point of an individual. This approach does not clearly express the core high and low density regions in the range of motion of the animal.

(2) Estimation methods based on nuclear density, including fixed nuclear methods and adaptive nuclear methods, can represent the probability density of the range of motion distribution of an animal or herd of animals. However, the conventional kernel density estimation method is based on euclidean distance, and cannot fully take into account the influence of terrain and obstacles on the distance between the space objects. In other words, these methods have difficulty in taking account of anisotropy of animal activities.

(3) The problem with the local convex hull approach is that it does not express fuzzy boundaries of the animal's range of motion. Nor the influence of terrain and obstacles.

Disclosure of Invention

The invention provides a home domain estimation technical method based on active learning and fuzzy mathematics. The home domain determining method considering the terrain and the obstacles under different geographic environments is realized. In the method, geographic information basic data is used for defining the coverage range of an obstacle, DEM data is used for determining terrain trend, GPS position data is combined with the obstacle data and the DEM data, and the numerical representation of the probability distribution of the animal movement range is realized through a fuzzy mathematical method. And extracting a core area and a non-core area (boundary area) of the active area by a threshold segmentation method, and further obtaining the home area description of the animal activity range.

The invention adopts active learning and fuzzy mathematics to construct a fuzzy home domain determination system. The fuzzy home domain determining system is composed of a data rasterization module, a cost distance calculating module, a possibility step calculating module and a home domain extracting module. Various geographic information data enter the system, and vector data are converted into a grid form through a rasterizing module. The cost distance calculation module combines the terrain data and the barrier data with the animal activity capability to obtain the cost distance from each animal position point to any point in space. The likelihood distribution calculation module is used for calculating the likelihood distribution of the animal moving in the area. The home domain extraction module divides the probability distribution data through a threshold value, so that a core area and a boundary area of the animal activity range are obtained, and the calculation of the home domain is realized.

Through the solution, the method for determining the activity range of the wild animal by using various geographic information data is realized, and the authenticity, reliability and accuracy of home domain estimation are improved.

In order to achieve the purpose, the invention discloses the following technical contents:

an animal home domain estimation method based on active learning is characterized in that: the method takes GPS point location data, DEM data and basic geographic information data as input data, and mainly comprises the following steps: the system comprises a data rasterization module, a cost distance calculation module, a possibility distribution calculation module and a home domain extraction module; the GPS point location data is tracking data of animal activities, the obstacles can be determined by basic geographic information data, and the obstacles of different animals are possibly different; the DEM data is used for determining the influence of terrain on the animal activity, and then a cost distance calculation module is used for calculating the cost distance from each GPS tracking point to any point in the animal activity area; calculating a fuzzy probability distribution of the animal moving in the area by using a probability distribution calculation module; the method comprises the following steps of extracting a core area and a non-core area (boundary area) of an animal activity range by utilizing a home area extraction module through determining a threshold value in advance, and realizing the calculation of a home area, wherein the process is as follows:

(1) data rasterization module

For different animals, different data types need to be selected according to the characteristics of the animals, and in the rasterization process, all data need to adopt a uniform coordinate system;

(2) cost distance calculation module

When considering obstacles and terrain in a natural environment, the sample points may be anisotropic, with an image size ofM×NM is the number of rows in the image and N is the number of columns in the image. The size of the adjacent window is 3 x 3, and then the cells in the image (the relief amplitude) are adjacent to one of its neighbors in FIG. 3aThe height difference therebetween is expressed as follows:

wherein

Is an arbitrary element in the DEM and,

is that

Any of the neighborhood of the image of the object,

are respectively

Calculating the cost distance from the original point to any point in the space through the following three steps;

step 1: setting the image size toM×N(ii) a Distance matrixD（M×N) For representing the distance of an arbitrary point from the source point, the initial distance is set to a very large value, e.g. a maximum of the integer type, if point (c: (b))m,n) Is the source point, set upD(m,n) Equal to 0;

step 2: scanning the matrix from the top left corner to the bottom right cornerDWhereini= 0, 1, 2, …, M–1，j= 0, 1, 2, …, N1 as shown in FIG. 3 (b). Distance matrixDThe updating is as follows:

（2）

whereinSIn order to be able to determine the resolution of the raster data,kparameters for adjusting the difficulty of movement; if it is not

Is an obstacle, and the animal cannot reach the pixel when moving, so that no calculation is performed on the pixel;

and step 3: scanning the matrix from the lower right corner to the upper left cornerDWhereini = M–1, M–2, …, 0，j = N–1, N2, …, 0, distance matrixDThe updating is as follows:

（3）

like if

If the obstacle is an obstacle, calculation is not needed;

kthe value depends on the living habits and the environment of the animal; notably, if the terrain is flat, the effect of the terrain factor may not be considered; in this case, it is preferable that the air conditioner,kset to 0;

in addition, different species of animals have different climbing abilities, when in uses(x) Is a function of the relief; therefore, the temperature of the molten metal is controlled,s(x) Is represented as follows:

（4）

wherein

Is a positionxThe degree of relief of the terrain at the location,

is a gaussian density function of the relief;

the larger the value, the higher the preference of the animal for the relief value and vice versa; the coefficient of influence of the final relief degree on the distance is expressed as follows:

(3) a likelihood distribution calculation module:

the motion of the animal is anisotropic, at which stage the Euclidean distance is replaced by a cost distance, and a triangular blur number is used to represent the range of influence of the sample pointDThe following were used:

whereindIs the distance from the target point to the source point,Dis the range of influence of the source point; the membership value of a point describes how much the point is affected by the sample point, so the greater the degree of influence, the greater the likelihood that the point belongs to the home domain.

Is provided withHIs a sampletIs obtained by conversion from the cost plane by equation (6), the complete ink droplet spreading plane is expressed as follows:

in the formula

Is a grid cell

The likelihood value of (d), abbreviated IDS value,Tis the number of source points;

is a source pointtTo unit

Then, normalizing the whole ink drop diffusion plane into a range of (0, 1), and determining the membership degree of the ink drop diffusion plane;

(4) a home domain extraction module;

respectively estimating threshold values of a core area and a boundary area by adopting 50% and 90% contour lines of membership degrees, and further extracting the core area and the boundary area of a home domain; sorting IDS values of all sampling points, and extracting IDS values of points at 50% of positions in a sorted queue as shown in the following formula;

then the area where the IDS value is greater than the threshold is the core area. The boundary region can be calculated in a similar way.

The invention further discloses application of the animal home domain estimation method based on active learning in effectively determining the activity range of wild animals and improving the authenticity of home domain estimation. The experimental results show that: the invention can determine the activity range of the wild animals by utilizing various geographic information data, and improves the authenticity, reliability and accuracy of home domain estimation.

Compared with the prior art, the animal home domain estimation method based on active learning disclosed by the invention has the positive effects that:

(1) the probability or likelihood of animal distribution can be naturally expressed;

(2) the range of the home area can effectively avoid obstacles such as cliffs, mountains, rivers and roads;

(3) the influence of the terrain can be considered, and the analysis result accords with the objective actual condition;

(4) the estimated home domain range error is minimal.

Drawings

FIG. 1 is a fuzzy home domain calculation process;

FIG. 2 is a graph of the cost distance from any point in grid space to point 1;

FIG. 3a cost distance calculation process; (a) target cell

8 neighborhoods; (b) from aboveThe following scanning order; (c) a bottom-up scan order;

FIG. 4 illustrates the area of influence of the source point 1 due to Euclidean distance and cost distance, respectively; wherein (a) the extent of influenceD；

(b) The Euclidean distance; (c) a cost distance;

FIG. 5 simulates a sika data set;

fig. 6 obstacles to sika activity are areas (marked a in the figure), ponds (marked B in the figure), motorway service areas (marked C in the figure), motorways (marked D in the figure) and farms (marked E in the figure) at an altitude of more than 1800 m;

FIG. 7 fuzzy membership of the simulated data;

FIG. 8 core and non-core regions.

Detailed Description

The invention is described below by means of specific embodiments. Unless otherwise specified, the technical means used in the present invention are well known to those skilled in the art. In addition, the embodiments should be considered illustrative, and not restrictive, of the scope of the invention, which is defined solely by the claims. It will be apparent to those skilled in the art that various changes and modifications can be made to these embodiments without departing from the spirit and scope of the invention.

Example 1

Taking various kinds of geographic information data (GPS point set, basic geographic information data and DEM data) as input data, wherein the GPS point position data set is tracking data of animal activities; the obstacles may be determined using basic geographic information data, and obstacles may be different for different animals; DEM data is used to determine the effect of terrain on animal activity. It should be noted that because of the diversity of animals, different data may be needed for different animals. The above three data are rasterized with the same resolution and the coordinate system is unified. Then, calculating the cost distance from each GPS tracking point to any point in the animal activity area by using a cost distance calculation module; calculating a fuzzy probability distribution of the animal moving in the area by using a probability distribution calculation module; the calculation of the home domain is realized by extracting the core region and the non-core region (boundary region) of the animal activity range by determining the threshold value in advance by using the home domain extraction module. The detailed flow is shown in figure 1:

1. data rasterization module

Since the present invention operates in grid space, all vector data that may be needed should be digitized into a grid format. Different data types need to be selected for different animals depending on the characteristics of the animal and therefore the system may require different data. In the rasterization process, a unified coordinate system is required to be adopted for all data.

2. Cost distance calculation module

When considering obstacles and terrain in a natural environment, the sample points may be anisotropic, with an image size ofM ×NThe neighboring window size is 3 x 3, then the height difference between the cell in the image (the relief amplitude) and its one neighborhood in fig. 3a is represented as follows:

wherein

Is an arbitrary element in the DEM and,

is that

Any of the neighborhood of the image of the object,

are respectively

The elevation of the point is calculated by the following three stepsCost distance of the intention;

step 1: setting the image size toM×N(ii) a Distance matrixD（M×N) For representing the distance of an arbitrary point from the source point, the initial distance is set to a very large value, e.g. a maximum of the integer type, if point (c: (b))m,n) Is the source point, set upD(m,n) Equal to 0;

step 2: scanning the matrix from the top left corner to the bottom right cornerDWhereini = 0, 1, 2, …, M–1，j = 0, 1, 2, …, N1 as shown in FIG. 3 (b). Distance matrixDThe updating is as follows:

（2）

whereinSIn order to be able to determine the resolution of the raster data,kparameters for adjusting the difficulty of movement; if it is not

Is an obstacle, and the animal cannot reach the pixel when moving, so that no calculation is performed on the pixel;

and step 3: scanning the matrix from the lower right corner to the upper left cornerDWhereini = M–1, M–2, …, 0，j = N–1, N2, …, 0, distance matrixDThe updating is as follows:

（3）

like if

If the obstacle is an obstacle, calculation is not needed;

kthe value depends on the living habits and the environment of the animal; notably, if the terrain is flat, the effect of the terrain factor may not be considered; in this case, it is preferable that the air conditioner,kset to 0;

in addition, different animal species differClimbing ability, at this times(x) Is a function of the relief; therefore, the temperature of the molten metal is controlled,s(x) Is represented as follows:

（4）

wherein

Is a positionxThe degree of relief of the terrain at the location,

is a gaussian density function of the relief;

the larger the value, the higher the preference of the animal for the relief value and vice versa; the coefficient of influence of the final relief degree on the distance is expressed as follows:

3. a likelihood distribution calculation module:

the motion of the animal is anisotropic, at which stage the Euclidean distance is replaced by a cost distance, and a triangular blur number is used to represent the range of influence of the sample pointDThe following were used:

whereindIs the distance from the target point to the source point,Dis the range of influence of the source point; the membership value of a point describes how much the point is affected by the sample point, so the greater the degree of influence, the greater the likelihood that the point belongs to the home domain.

Is provided withHIs a sampletIs obtained by conversion from the cost plane by equation (6), the complete ink droplet spreading plane is expressed as follows:

whereinTIs the number of source points;

is a source pointtTo unit

Then, normalizing the whole ink drop diffusion plane into a range of (0, 1), and determining the membership degree of the ink drop diffusion plane;

4. a home domain extraction module:

respectively estimating threshold values of a core area and a boundary area by adopting 50% and 90% contour lines of membership degrees, and further extracting the core area and the boundary area of a home domain; sorting IDS values of all sampling points, and extracting IDS values of points at 50% of positions in a sorted queue as shown in the following formula;

then the area where the IDS value is greater than the threshold is the core area. The boundary region can be calculated in a similar way.

Example 2

Simulation data set as shown in fig. 5, the simulated animal was a sika that liked to live at the edge of forests and lawns, but did not like to live in dense forests or shrubs; in addition, they prefer areas with little human interference, open space, and abundant water sources. The spatial resolution of DEM data is 25m and the elevation is between 689m and 2129 m. In this example, areas with an altitude above 1800m (marked a in fig. 6), ponds (marked B in fig. 6), motorway service areas (marked C in fig. 6), motorways (marked D in fig. 6) and farms (marked E in fig. 6) are considered as obstacles to sika activity. Since animals prefer to live on flat terrain rather than rugged steep terrain, the relief from the DEM is used to measure the cost distance in this area. By using the method of the invention, the fuzzy membership of the home domain of the simulation data set is obtained, as shown in FIG. 7; the extracted home domain core area and non-core area (border area) are shown in fig. 8.

And (4) conclusion:

the results of the present invention were compared with those of the core density estimation method and the local convex hull method, and the results are shown in table 1. From the area comparison, the partial convex hull method 2: (k=16) results in a minimum core area, but a very large non-core area. Graphically, the kernel density estimation method and the method of the invention can describe the gradual transition characteristics of the home domain, the graph is very broken, and the kernel density estimation method and the local convex hull method cannot avoid the obstacles.

The method of the invention can completely avoid obstacles, can describe gradually transitional home domain characteristics and obtains the minimum total coverage area, so the method of the invention is superior to the two methods.

Table 1. the method, kernel density estimation method and local convex hull method of the present invention estimate the area of the core region and non-core region of the home domain of the simulation dataset:

Claims (1)

1. an animal home domain estimation method based on active learning is characterized in that: the method takes GPS point location data, DEM data and basic geographic information data as input data, and mainly comprises the following steps: the system comprises a data rasterization module, a cost distance calculation module, a possibility distribution calculation module and a home domain extraction module; the GPS point location data is tracking data of animal activities, the obstacles can be determined by basic geographic information data, and the obstacles of different animals are possibly different; the DEM data is used for determining the influence of terrain on the animal activity, and then a cost distance calculation module is used for calculating the cost distance from each GPS tracking point to any point in the animal activity area; calculating a fuzzy probability distribution of the animal moving in the area by using a probability distribution calculation module; the method comprises the following steps of extracting a core area and a boundary area of an animal activity range by utilizing a home area extraction module through determining a threshold value in advance, and realizing the calculation of a home area, wherein the process is as follows:

(1) data rasterization module

For different animals, different data types need to be selected according to the characteristics of the animals, and in the rasterization process, all data need to adopt a uniform coordinate system;

(2) cost distance calculation module

When considering obstacles and terrain in a natural environment, the sample points may be anisotropic, with an image size ofM ×NM is the number of rows in the image and N is the number of columns in the image.

The neighbor window size is 3 x 3, and then the height difference between the cell and the neighborhood in the image is represented as follows:

（1）

wherein

Is an arbitrary element in the DEM and,

is that

Any of the neighborhood of the image of the object,

are respectively

Calculating the cost distance from the original point to any point in the space through the following three steps;

step 1: setting the image size toM×N(ii) a Distance momentMatrix ofD（M×N) For representing the distance of an arbitrary point from the source point, the initial distance is set to a very large value, e.g. a maximum of the integer type, if point (c: (b))m,n) Is the source point, set upD(m,n) Equal to 0;

step 2: scanning the matrix from the top left corner to the bottom right cornerDWhereini= 0, 1, 2, …, M–1，j = 0, 1, 2, …, N-1, distance matrixDThe updating is as follows:

(2)

whereinSIn order to be able to determine the resolution of the raster data,kparameters for adjusting the difficulty of movement; if it is not

Is an obstacle, and the animal cannot reach the pixel when moving, so that no calculation is performed on the pixel;

and step 3: scanning the matrix from the lower right corner to the upper left cornerDWhereini= M–1, M–2, …, 0，j= N–1, N2, …, 0, distance matrixDThe updating is as follows:

（3）

if it is not

If the obstacle is an obstacle, calculation is not needed;

kthe value depends on the living habits and the environment of the animal; notably, if the terrain is flat, the effect of the terrain factor may not be considered; in this case, it is preferable that the air conditioner,kset to 0;

in addition, different species of animals have different climbing abilities, when in uses(x) Is a function of the relief; therefore, the temperature of the molten metal is controlled,s(x) Is represented as follows:

（4）

wherein

Is a positionxThe degree of relief of the terrain at the location,

is a gaussian density function of the relief;

the larger the value, the higher the preference of the animal for the relief value and vice versa; the coefficient of influence of the final relief degree on the distance is expressed as follows:

(3) a likelihood distribution calculation module:

the motion of the animal is anisotropic, at which stage the Euclidean distance is replaced by a cost distance, and a triangular blur number is used to represent the range of influence of the sample pointDThe following were used:

whereindIs the distance from the target point to the source point,Dis the range of influence of the source point; the membership value of a point describes the degree of influence of the sample point on the point, so the greater the degree of influence, the greater the probability that the point belongs to the home domain;

is provided withHIs a sampletIs obtained by conversion from the cost plane by equation (6), the complete ink droplet spreading plane is expressed as follows:

in the formula

Is a grid cell

The likelihood value of (d), abbreviated IDS value,Tis the number of source points;

is a source pointtTo unit

Then, normalizing the whole ink drop diffusion plane into a range of (0, 1), and determining the membership degree of the ink drop diffusion plane;

(4) a home domain extraction module;

respectively estimating threshold values of a core area and a boundary area by adopting 50% and 90% contour lines of membership degrees, and further extracting the core area and the boundary area of a home domain; sorting IDS values of all sampling points, and extracting IDS values of points at 50% of positions in a sorted queue as shown in the following formula;

then the area where the IDS value is greater than the threshold is the core area.

Images