An Inland Aquaculture Pond Extraction Method Based on Angle Constraint and Double-Neighborhood Statistics

Abstract

The present invention provides a method for extracting inland aquaculture ponds based on angle constraint and double-neighborhood statistics, including: obtaining a Landsat-8 image, preprocessing the Landsat-8 image, and extracting a water body area from the preprocessed image; connecting broken rivers in the extracted water body area according to the similarity of the tangent angle directions in the local space between the broken rivers, so as to obtain a river area; extracting aquaculture ponds near the river area; extracting aquaculture ponds around lakes and reservoirs; discriminating the aquaculture areas in the river area, the aquaculture areas around lakes and reservoirs from three aspects of direction similarity, geometric shape similarity and area singular value, and finally obtaining the extraction result of inland aquaculture ponds. The present invention extracts inland aquaculture ponds by combining the double-angle curvature feature with the direction and double-neighborhood characteristics of the spatial distribution of inland aquaculture ponds, improving the extraction efficiency and accuracy.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and more particularly, to a method for extracting inland aquaculture ponds based on angle constraints and double-neighborhood statistics. Background Art

[0002] The output of inland aquaculture accounts for 59% of the total output of aquaculture, and it is still one of the main ways of aquaculture at present, which can provide high-quality fishery products for our country and bring huge economic benefits. However, due to the breeding characteristics and requirements of economic products such as fish and shrimps, the vast majority of inland aquaculture ponds are distributed in or around rivers, reservoirs, and lakes. This distribution characteristic also brings many adverse effects, such as the destruction of traditional habitats of aquatic animals and plants, the deterioration of fresh water quality, and the illegal construction of aquaculture ponds. Therefore, it is necessary to timely extract the areas of inland aquaculture ponds for convenient monitoring. In the early stage, the method of on-site investigation was mostly used to extract inland aquaculture ponds. However, the disadvantage of this method is that it is time-consuming and laborious, and it is difficult to carry out large-scale detection and mapping work on aquaculture ponds. With the development of satellite remote sensing technology, it has gradually replaced the traditional method of on-site investigation with its advantages such as fast speed and wide coverage area, and thus has been favored by many scientific researchers.

[0003] Due to different forms of inland aquaculture, inland aquaculture ponds are divided into net cage aquaculture and inland pond aquaculture. At present, most studies focus on the extraction of net cage aquaculture areas in inland lakes. Representative extraction methods include the method based on texture features and principal component analysis, and the method based on the combination of spectral information, texture information, and long time series. The above studies have all achieved good extraction results. However, there are relatively few studies on inland pond aquaculture. Some scholars only selected the Liangzi Lake area in Landsat 8 images and GF-1 images, and used the more regular boundary of artificial water bodies to distinguish inland aquaculture ponds and lakes around the lakes.

[0004] However, in the medium-resolution images, the water body types and the location distributions of aquaculture ponds in areas other than lakes are complex, making the extraction of inland aquaculture ponds challenging in three aspects. First, the geometric shape features and directions of inland aquaculture ponds are similar to those of nearby truncated rivers, making it difficult to distinguish them. This is because the image resolution is low and there are bridge intervals, resulting in some rivers being truncated. At this time, the geometric shape of the truncated rivers is more regular. Worse still, the truncated rivers are adjacent to the inland aquaculture ponds around them, resulting in similarity in the connectivity domain directions of the two. Second, there are many small non-aquaculture ponds in the inland area, which are extremely widely distributed. Their geometric shape features are similar to those of some inland aquaculture ponds, further increasing the difficulty of extracting inland aquaculture ponds. Finally, artificial reservoirs and inland aquaculture ponds are difficult to distinguish due to similar geometric features. The reason is that both the reservoir and the inland aquaculture pond are artificial water bodies, and they have high similarity in geometric features such as shape, resulting in artificial reservoirs being easily misclassified as inland aquaculture ponds in medium-resolution images. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention provides a method for extracting inland aquaculture ponds based on angle constraint and double-neighborhood statistics. The present invention uses a method that combines the double-angle curvature feature with the direction and double-neighborhood characteristics of the spatial distribution of inland aquaculture ponds to extract inland aquaculture ponds, and there is at least a 5% improvement in both kappa coefficient and IOU evaluation indicators.

[0006] The technical means adopted by the present invention are as follows:

[0007] A method for extracting inland aquaculture ponds based on angle constraint and double-neighborhood statistics, comprising:

[0008] Obtain Landsat-8 images, preprocess the Landsat-8 images, and extract the water body area from the preprocessed images;

[0009] According to the similarity of the tangent angle directions in the local space between the fractured rivers, connect the fractured rivers in the extracted water body area to obtain the river area;

[0010] Extract the aquaculture ponds near the river area;

[0011] Extract the aquaculture ponds around lakes and reservoirs;

[0012] Discriminate the aquaculture areas in the river area, the aquaculture areas around lakes and reservoirs from three aspects: direction similarity, geometric shape similarity, and area singular value, and finally obtain the extraction result of inland aquaculture ponds.

[0013] Further, preprocess the Landsat-8 image, including: performing radiometric calibration and atmospheric correction on the Landsat-8 image.

[0014] Further, according to the similarity of the tangent angle directions in the local space between the broken rivers, connect the broken rivers in the extracted water body area, including:

[0015] Obtain the endpoints of the rivers to be connected, select a local neighborhood within the endpoints, and obtain another point on the river within the local neighborhood;

[0016] Use the average slope of the local neighborhood space and the average of the tangent slopes at the endpoints to characterize the direction of the truncated river, so as to obtain the tangent angle representing the river flow direction;

[0017] When the tangent angles of the river flow directions at the two endpoints of the river truncation satisfy a preset relationship, connect the two endpoints.

[0018] Further, according to the similarity of the tangent angle directions in the local space between the broken rivers, connect the broken rivers in the extracted water body area, including:

[0019] Make tangents at the endpoints of the two rivers to be connected, obtain the intersection point of the two tangents, use the two river endpoints and this intersection point as endpoints, and construct a triangular area to be connected;

[0020] Connect the rivers according to the number of broken river segments in the area to be connected, including:

[0021] When the number of river segments is not less than 1, determine the connection order according to the distance metric of each river segment,

[0022] When the number of river segments is 0, use the midpoint of the triangle median line as the transition point for river connection to connect the rivers.

[0023] Further, extract the aquaculture ponds near the river area, including:

[0024] Obtain the curvature of each pixel along the water body boundary;

[0025] Construct a curvature feature based on the curvature of each pixel, and the curvature feature is used to describe the regularity of the water body contour;

[0026] According to the direction characteristics of the aquaculture ponds and the river direction, construct an extraction feature, and the extraction feature is used to describe the spatial relationship between the aquaculture ponds and the river;

[0027] Compare the curvature feature and the extraction feature extracted from the image with the preset curvature threshold and extraction threshold, and extract the area that meets the requirements as the extracted river-type aquaculture pond area.

[0028] Further, for the extraction of aquaculture ponds around lakes and reservoirs, it includes:

[0029] Obtain the curvature of each pixel along the water body boundary;

[0030] Construct a curvature feature based on the curvature of each pixel, and the curvature feature is used to describe the regularity of the water body contour;

[0031] Judge that if a connected domain with a curvature feature meeting the preset requirements contains another connected domain with the same curvature feature meeting the preset requirements in its local space, then judge that the current connected domain is a lake aquaculture area.

[0032] Compared with the prior art, the present invention has the following advantages:

[0033] This application proposes an inland aquaculture pond extraction algorithm based on angle constraint and double-neighborhood statistics. This algorithm uses a method that combines double-angle curvature features with the direction and double-neighborhood characteristics of the spatial distribution of inland aquaculture ponds to extract inland aquaculture ponds. Through experiments, the extraction results of the method in this application and the comparison method for inland aquaculture ponds are compared, and the performance of the two methods in different regions and different water body types is analyzed. In order to verify the effectiveness of the algorithm in this application, two evaluation indicators, namely the kappa coefficient and the IOU, are used to quantify the experimental results. The experiments show that compared with the comparison algorithm, this application has at least a 5% improvement in both evaluation indicators. Description of the Drawings

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0035] Figure 1 It is a flowchart of a method for extracting inland aquaculture ponds based on angle constraint and double-neighborhood statistics according to the present invention.

[0036] Figure 2 It is a schematic diagram of the river flow direction in the embodiment.

[0037] Figure 3 It is an 8-neighborhood schematic diagram of the river end point in the embodiment.

[0038] Figure 4 It is a schematic diagram of the river end point in the embodiment.

[0039] Figure 5 It is a schematic diagram for solving the approximate direction angle in the local neighborhood space in the embodiment.

[0040] Figure 6 Schematic diagram of river distribution in the embodiment, where (a) is the satellite image of Google Maps, (b) is the true river map, (c) is the schematic diagram of River A, and (d) is the schematic diagram of River B.

[0041] Figure 7 Schematic diagram of river connection in the embodiment, where (a) is the schematic diagram of the connection of River A and (b) is the schematic diagram of the connection of River B.

[0042] Figure 8 Distribution map of inland aquaculture ponds in the embodiment, where (a) is the satellite image of Google Maps, (b) is the distribution map of rivers and aquaculture ponds, and (c) is the enlarged view of the yellow rectangular frame area.

[0043] Figure 9 Schematic diagram of directions in pixels in the embodiment, where (a) is the chain code direction map, (b) is where θ(Pi) is 0, and (c) is where θ(Pi) is π / 2.

[0044] Figure 10 Schematic diagram of the directions of rivers and aquaculture ponds in the embodiment.

[0045] Figure 11 Schematic diagram of the spatial characteristics of aquaculture ponds in the embodiment.

[0046] Figure 12 Schematic diagram of the directions of aquaculture ponds in a local area in the embodiment.

[0047] Figure 13 Distribution map of the positions of artificial reservoirs and aquaculture ponds under satellite images in the embodiment, where (a) is the satellite image of Google Maps and (b) is the distribution map of the positions of artificial reservoirs and aquaculture ponds.

[0048] Figure 14 Experimental results of Landsat-8 images in Tianjin in the embodiment, where (a) is the RGB false color image, (b) is the true map of the aquaculture pond area, (c) is the extraction result of WRIP aquaculture ponds, and (d) is the extraction result of aquaculture ponds in this article.

[0049] Figure 15 Experimental results of Landsat-8 images in Zhongshan in the embodiment, where (a) is the RGB false color image, (b) is the true map of the aquaculture pond area, (c) is the extraction result of WRIP aquaculture ponds, and (d) is the extraction result of aquaculture ponds in this article. Detailed implementation manners

[0050] To enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0051] The present invention provides a method for extracting inland aquaculture ponds based on angle constraint and double-neighborhood statistics, which mainly includes five steps, as Figure 1 shown. The first step is the preprocessing and water body extraction of Landsat-8 images, which mainly includes radiometric calibration, atmospheric correction, and the Modified Normalized Difference Water Index (MNDWI) and Hessian matrix. The second step is river connection, and the main idea is to utilize the similarity of the tangent angle direction in the local space between broken rivers to complete the connection of broken rivers. The third step is to extract the aquaculture ponds near the rivers, the fourth step is to extract the aquaculture ponds around lakes and reservoirs, and the fifth step is to discriminate the inland aquaculture ponds extracted in the third and fourth steps from three aspects: direction similarity, geometric shape similarity, and area singular value, and finally obtain the extraction result of the inland aquaculture ponds. The following will specifically describe each step.

[0052] 1. Preprocessing and Water Body Extraction

[0053] To obtain the spectral reflectance of Landsat-8 images, it is necessary to preprocess the original Landsat-8 images. The preprocessing includes radiometric calibration and atmospheric correction. Radiometric calibration is to correct the radiation distortion to convert the original digital quantization value into a radiance value, so as to eliminate the error caused by the sensor itself and determine the accurate radiation value at the sensor entrance. Atmospheric correction is to eliminate the error caused by atmospheric scattering, absorption, and reflection, so as to convert the radiance value into the actual surface reflectance. In ENVI5.3, the radiometric calibration tool (Radiometric Calibration) is generally used to complete the radiometric calibration, and the atmospheric correction is also completed under the Radiometric Calibration toolbox, mainly requiring the operator to participate in some parameter settings, such as selecting a suitable atmospheric model according to the image imaging time and dimension information, etc. Next, the water content of each pixel in the image is calculated through MNDWI to obtain the binary image of the water body of the remote sensing image. In order to keep more detailed information on the edge of the water body and show a better image effect, the Hessian matrix is also used to enhance the edge of the image.

[0054] 2. River connection strategy based on endpoint triangles

[0055] Inland rivers contain rich fresh water resources. Some inland aquaculture ponds are usually distributed on both sides of the rivers. Therefore, the direct idea of detecting inland aquaculture ponds is to detect inland rivers. Rivers usually show a meandering characteristic due to topographical reasons. So, in this application, the main rivers are extracted using existing technologies. Since the resolution of Landsat-8 images is 30 meters, the width of most rivers is only a few pixel points. Also, due to the narrowing of the river channels at the tributaries and the intervals between bridges, some rivers are truncated, making it difficult for them to maintain integrity. To effectively reduce the interference of rivers on the extraction of inland aquaculture ponds, avoid misclassification of truncated rivers and surrounding aquaculture ponds, and for subsequent work to locate inland aquaculture ponds relying on complete rivers, it is necessary to connect the broken rivers.

[0056] As Figure 2 shown, the curve represents the river, and the arrow and line segment represent the direction and local neighborhood space within the local neighborhood of the river respectively. Obviously, the river has the characteristic of unchanged flow direction within the local neighborhood. Therefore, this application gives a river connection strategy based on local tangent angle similarity based on this feature. To determine the local flow direction of the rivers to be connected, it is necessary to select the length of the local neighborhood at the truncated part of the rivers to be connected. Considering the similarity of local directions, this application compromises and selects a river line segment with a length of 10 pixel points as the local object. To determine the local object of the river, first, its endpoints need to be clarified. The river is represented as a curve consistent with the connectivity and topological structure of the original shape, and its pixels are generally 1. In a binary image, if a pixel point with a pixel value of 1 has only 1 point with a pixel value of 1 other than itself within its 8-neighborhood, then this point is an endpoint of a curve. The schematic diagram is as Figure 3 shown, and the pixel point is the endpoint of the line segment to which this point belongs. After obtaining an endpoint of the truncated river through the above operations, it is necessary to obtain another endpoint within the local neighborhood of this river. As Figure 4 shown, the curve represents the river, the points represent the endpoints of the river, and the box-marked area contains two rivers to be connected. Given an endpoint P of the river and a local neighborhood length of 10, another point, point Q, in the local neighborhood space can be obtained.

[0057] After determining the local neighborhood space of the river, to more accurately represent the local flow direction of the river, the average slope of the local neighborhood space is used to represent the river direction. However, although the length of the local neighborhood space is only 10 pixels, inevitably, the river line segment within the local neighborhood space will bend, and then the difference between the average slope of the local neighborhood space and the actual flow direction of the river will increase. To illustrate this problem, the schematic diagram is as Figure 5As shown in the figure, P and Q respectively represent two end points in the local neighborhood space, and point N is the adjacent pixel of point P on the river skeleton line. Since the river is curved in the local neighborhood space, the average slope θ of the local neighborhood space PQ differs greatly from the true flow direction θ PN of the cyan river. To reduce this difference and thus characterize the true flow direction of the truncated river, the present application uses the average of the average slope of the local neighborhood space and the tangent slope at the end points to characterize the direction of the truncated river, that is Figure 5 the average of θ PQ and θ PN , denoted as k, which can be expressed as:

[0058]

[0059] where P X , Q X , N X and P Y , Q Y , N Y are the horizontal and vertical coordinates of points P, Q, and N in the image respectively. Therefore, the tangent angle α representing the river flow direction can be obtained through the local neighborhood space, which can be expressed as:

[0060] α = |arctank - arctank0| (2)

[0061] where k0 represents the slope of the X-axis. To connect the known river and the truncated river, using the similarity of the river flow direction within the local neighborhood range, the approximate direction angle relationship between the two rivers needs to satisfy:

[0062] |α1 - α2| ≤ T (3)

[0063] where α1 and α2 are the approximate direction angles of two adjacent rivers to be connected respectively. Due to pixel reasons, it is difficult to ensure absolute parallelism in the mathematical sense, so the threshold T in the present application is generally taken as 10 degrees. When the two rivers satisfy the above relationship, the end points of the rivers are connected to form a complete river. However, when the river is in the narrow channel position of the tributary or the river is truncated in multiple places due to sediment deposition, bridge spacing, etc., there are two cases at this time, namely the truncated river satisfies the extraction method of the existing technology and the truncated river does not satisfy this extraction method, as Figure 6 shown Figure 6 (a) is the river under the satellite image, and the above two cases are respectively marked as A and B with blue frames in Figure 6 (b), where the schematic diagrams of river extraction are respectively as Figure 6 (c) and Figure 6 (d).

[0064] In Figure 6(c), cyan and green respectively represent the extracted rivers and the truncated small river segments. At this time, the tangent angle relationship between the two extracted rivers obviously does not meet the threshold discrimination in formula (3), resulting in the inability to connect two adjacent river segments. While in Figure 6 (d), only the river represented by cyan exists. Due to the low surface water content, there is no situation like the green river segment in Figure 6 (c) between the two rivers. To connect the rivers in the above two situations and reduce the deviation of estimating the actual flow direction of the river, tangents are made at the endpoints of the two rivers to be connected, and the two tangents intersect at point M. These three endpoints form a triangular connection area to be connected. To determine the breakage situation of the rivers to be connected and thus connect them in different situations, the number of truncated river segments within the triangular area is denoted as R t , which can be expressed as:

[0065]

[0066] where r is the mark of the truncated river segment existing within the triangular area. When i is not less than 1, there is no less than 1 truncated river segment within the triangular area at this time, that is, the situation in Figure 6 (c). To connect the truncated rivers in sequence, the present application uses distance measurement to determine the connection order. Select one of the two rivers to be connected and mark its endpoint as the starting point P(x, y). Then, the set of distances from point P to the endpoints of all truncated rivers within the triangular area is arranged in ascending order, denoted as U, which can be expressed as:

[0067]

[0068] where j represents the jth endpoint within the triangular area. Starting from the starting point P(x, y), connect them in sequence according to the distances in the above set, and continuously update point P during the connection process until the point coincides with the endpoint on the other river to be connected, indicating the completion of the connection. The connection schematic diagram is as shown in Figure 7 (a), and the black line segment represents the connection line.

[0069] When i is equal to 0, that is, there is no truncated river segment within the triangular area, that is, the situation in Figure 6 (d). To keep the meandering flow direction of the river unchanged as much as possible after connection, the application uses the midpoint of the triangle median line as the transition point for river connection. As shown in Figure 7 (b), the midpoint of the triangle median line corresponding to point M is M0. Connect the endpoints of the two rivers in sequence and the black connection line passes through M0 to obtain the complete river.

[0070] It should be noted that in order to prevent the incorrect connection of rivers with other interfering water bodies, the connection strategy needs to be restricted by the river width. This restriction is that the river width changes little in a local area and the width of one river is generally not less than half of the width of another river. Therefore, the width of the river to be connected should satisfy:

[0071] W r1 ≤2W r2 (6)

[0072] where W r1 and W r2 represent the widths at the endpoints of two rivers to be connected respectively.

[0073] 3. Extraction of river-type aquaculture ponds based on double-angle curvature and local direction similarity

[0074] After the river connection strategy in the previous section, the rivers disconnected due to low pixel resolution and bridge intervals are connected. There are two reasons for the connection: one is that the broken rivers will interfere with the extraction of inland aquaculture ponds, and the other is that some inland aquaculture ponds have special geographical location relationships with rivers, and detecting rivers is beneficial to the extraction of these aquaculture ponds. To illustrate this, take Figure 8 as an example. Figure 8 (a) Inland aquaculture ponds are usually built along rivers because the flowing river water can provide fresh food and a suitable living environment for aquaculture products. At the same time, due to economic costs, inland aquaculture ponds are often built within the river water area and as close to the river channel as possible. As shown in Figure 8 (b), the cyan diagonal area is the main channel of the river, and the inland aquaculture ponds distributed along the river are within the red marked frames. In order to reduce the water exchange cost, the distribution direction of a single pond is usually perpendicular or parallel to the river direction corresponding to its location. Similarly, in order to reduce the construction cost and facilitate manual management, the ponds are usually built in rectangles, as shown in Figure 8 (c). Although the aquaculture ponds are all regular rectangles, due to the water exchange ditches between the ponds, there are situations where multiple aquaculture ponds are connected together in the binary water body map. Therefore, the aquaculture ponds may not all remain as single connected regions in the Landsat-8 image. However, aquaculture ponds usually have regular boundaries, and the dikes are also straight lines, which are very different from the smooth and curved surface of natural water bodies. Based on the direction distribution and regular boundary characteristics of inland aquaculture ponds, this application presents an extraction algorithm based on the improved curvature feature and local direction similarity of angles to solve the problem of extracting inland aquaculture ponds along rivers.

[0075] First, the regular boundaries of aquaculture ponds are usually straight lines, and curvature can exactly describe the degree of curve bending. Since digital images are discrete representations of real-world objects, the curvature of each pixel of the contour line is approximately calculated using the angular change along the curved path. Therefore, in the actual pixel operations, the distance between the center points of the pixels is approximated to the arc length. Thus, the curvature of each pixel along the water body boundary can be calculated. Denote the curvature as γ(t), which can be expressed as:

[0076]

[0077] where ds is the change in arc length, and is the directional change of the angle relative to the arc length. Different curvatures represent different degrees of boundary bending at the pixel point. In binary images, not only zero curvature but also the curvature value description of the angular change at the rectangular corners characterizes the regular shape. In Figure 9 (a), the chain code is used to represent the eight directions of a pixel point in the image, Figure 9 (b) shows the case where the directional change is 0°, that is, the curvature is 0, representing the straight boundary of the aquaculture pond, and (c) shows the case where the directional change is π / 2, representing the rectangular right angle of the aquaculture pond. Among them, P i represents the current pixel point for calculating the curvature, and the direction of P i is denoted as θ(P i ), which can be expressed as:

[0078]

[0079] where and are the chain code values from P i to P i+1 and from P i-1 to P i respectively. Using the above formula, the curvature value of the pixel point can be obtained from the chain code direction. Based on this, the present application constructs a feature using angular curvature, denoted as C, which can be expressed as:

[0080]

[0081] where S γ＝0 and S γ＝π2 are the sums of the pixel points with curvature values of 0 and π / 2 on the boundary respectively, and S sumIt is the sum of all pixel points on the boundary. Using this formula, the percentages of the two regular parts, namely straight lines and right angles, in the water body boundary can be measured, and the regularity degree of the water body contour is described. The larger the C, the more the water body boundary tends to be a regular polygon. The boundary of an aquaculture pond is artificially constructed, and its boundary is generally a straight line segment and most of the corners are right angles. Therefore, formula (9) can be used to measure the regularity degree of the aquaculture pond.

[0082] In addition, if the dam is not obvious due to pixel reasons and multiple aquaculture ponds are connected together, then the number of right angles will also be more than 4 of a single aquaculture pond. Therefore, compared with other water bodies, the proportion of points with curvature of 0 and π / 2 in aquaculture ponds is higher. This application believes that a water body with C not less than 0.8 is an artificially constructed aquaculture pond.

[0083] Secondly, the distribution direction of aquaculture ponds and the river direction have regularity, that is, the direction of aquaculture ponds and the river direction have an approximately parallel or perpendicular spatial relationship. Therefore, features are extracted based on the local direction similarity between the river and aquaculture ponds. As Figure 10 shown, where the cyan curve and the black solid rectangle represent the river skeleton line and the connected domain of a single inland aquaculture pond respectively, and the yellow arrow represents the shortest distance from the centroid of the aquaculture pond connected domain to the river, which is the red point mapped onto the river. The green and red arrows represent the long axis direction of the aquaculture pond connected domain and the tangent direction of the red point in the river respectively, denoted as θ P and θ R , in order to construct the direction similarity between the two, the connected domain direction angle θ is defined as the intersection angle between the long axis of the ellipse with the same standard second-order central moment as the region and the X-axis. By calculating the connected domain direction angle, the relationship between the directions of the river and aquaculture ponds can be determined. The perpendicular line segment from the centroid of the aquaculture pond connected domain to the river intersects the river at the red point as shown in Figure 10 the black dashed box A, and the tangent direction of this point represents the river flow direction here. If it is similar to the direction of the aquaculture pond connected domain, it can be defined as:

[0084] |θ R -θ P |≤T (10)

[0085] where θ P and θ R are the connected domain direction of the aquaculture pond and the river direction respectively. Formula (10) describes the parallel similarity relationship between the two. However, when the distribution of the aquaculture pond and river directions is as shown in Figure 10 the black dashed box B, at this time the relationship between θ P and θ R satisfies:

[0086]

[0087] Therefore, when the direction of the connected domain of the aquaculture ponds near the river and the local tangent direction of the river satisfy formula (10) or formula (11), it conforms to the direction feature of aquaculture pond extraction. Finally, morphological operations and minimum bounding rectangle marking are performed on the connected domains that meet the improved curvature feature and local direction similarity of the angle, so as to obtain the rectangular areas of inland aquaculture ponds distributed along the river, while removing the river. The reason for the above operations is that inland aquaculture ponds are characterized by dense distribution, and at the same time, to avoid the situation that some aquaculture ponds are omitted because they do not meet the above features due to non-standard construction. Through the above analysis of the extraction algorithm for river-type inland aquaculture ponds, the rectangular areas of aquaculture ponds distributed along the river can be obtained. In order to prevent interference with the extraction of the remaining aquaculture ponds, the rectangular areas of the extracted aquaculture ponds are masked.

[0088] 4. Extraction of aquaculture ponds in lakes and reservoirs based on double-neighborhood statistics

[0089] Among inland areas, in addition to rivers containing rich fresh water resources, lakes are the next. Lakes have moderate water depths, are rich in aquatic biological resources, and have high biodiversity, making them very suitable for the cultivation of economic aquatic products. Therefore, there are often large areas of aquaculture ponds near lakes. Although the two are geographically adjacent, since lakes are natural water bodies, the lake surface is meandering and irregular to adapt to its local terrain. Obviously, the proportion of zero-curvature and π / 2 pixel points in lakes is lower, and formula (9) can be used to separate lakes from artificially constructed aquaculture ponds. However, there are still many small water bodies with small connected domain areas in inland areas, most of which are composed of small ponds, which interfere with the extraction of aquaculture ponds. At the same time, the aquaculture ponds near lakes may not all meet the improved curvature feature based on angles due to irregular construction.

[0090] To address the above problems, the present application uses the spatial feature of dense distribution of aquaculture ponds for extraction. The spatial feature is: if a connected domain that meets the improved curvature feature contains another connected domain that also meets the improved curvature feature in its local space, then it is considered that the connected domain conforms to the spatial feature of aquaculture ponds. As Figure 11 shown, the white grid represents the aquaculture pond, and the red dot is the centroid of a single connected domain that meets the improved curvature feature. Taking the diameter of the circumscribed circle of the connected domain as the radius R for local space search and its centroid as the center, the local neighborhood of the connected domain can be obtained. Except for the gray connected domain itself, if the number of aquaculture ponds in its local neighborhood is denoted as LP, it can be expressed as:

[0091]

[0092] where Q in and Q outThey represent the connected domains that are completely inside the circle and partially inside the circle, respectively, and the connected domains satisfy the improved curvature characteristics, otherwise they are not counted. The spatial distribution of aquaculture ponds is very compact, and there is a narrow dam between each aquaculture pond, so the LP of aquaculture ponds is usually not less than 1. In contrast, the ponds are surrounded by land, and there are generally no connected domains that meet the improved curvature characteristics in the local space, so their LP is often 0. Similarly, in order to prevent the omission of some aquaculture ponds that do not meet the above conditions, the connected domains that meet the above two characteristics at the same time are subjected to morphological operations and minimum circumscribed rectangle marking to obtain the rectangular area containing the water body to be identified.

[0093] 5. Identification of inland breeding ponds

[0094] Through the extraction of inland breeding ponds in the previous section, a rectangular area containing the inland breeding ponds can be obtained. This area includes two types of inland breeding pond water bodies that meet the above extraction criteria and water bodies that do not meet the above extraction criteria. Among them, the water bodies that do not meet the above extraction criteria are the water bodies to be identified. In order to further extract inland breeding ponds, it is necessary to screen the water bodies to be identified in the rectangular area to obtain all the breeding pond water bodies. Since the distribution of breeding ponds is clustered and the differences in area and direction between breeding ponds are small, this application uses the similarity between the extracted inland breeding ponds and the water bodies to be identified to construct breeding pond identification criteria. First, the direction of a single breeding pond and the rectangular area in which it is located is similar, such as Figure 12 As shown in the figure, the gray area in the rectangular coordinate system represents the area after the morphological operation of the breeding pond, and θLA and θS represent the overall direction of the area and the direction of a single breeding pond, respectively. The direction of a single pond and its region satisfies the mathematical relationship of being approximately perpendicular or parallel, so the direction similarity measurement of the water body to be identified covered by the local area satisfies the relationship:

[0095] |θ LA -θ S |≤T (13) or

[0096]

[0097] Wherein, T represents the threshold value, and its value remains the same as above.

[0098] Secondly, the aquaculture ponds are similar in shape and most of them are regular geometric shapes. When comparing artificially constructed aquaculture ponds and natural water bodies of the same size, it is obvious that the area of artificially constructed water bodies accounts for a larger proportion. Therefore, this feature is used to construct a shape similarity feature for aquaculture pond discrimination. If the shape feature is recorded as Ip, it can be expressed as:

[0099]

[0100] Among them, S and SREC They are respectively the area of the water body connection domain in the rectangular area and the area of its minimum circumscribed rectangle, I p which measures the fitting degree between the contour of the water body and the rectangle. Although the existing technology believes that a large number of aquaculture ponds usually satisfy the condition that the area is less than 1 km 2 , I p is not less than 0.8, but in order to improve the automation of the algorithm in this chapter, this application uses the statistic of the shape index of the extracted aquaculture ponds as the threshold. Statistically calculate the mean value of the extracted aquaculture ponds I p in a rectangular area as the threshold of the water body characteristics to be discriminated I p , and the relationship they satisfy is:

[0101]

[0102] where I p (d) and I p (q) are respectively the I p values of the water body to be discriminated and the water body of the extracted inland aquaculture ponds. The meaning of this formula is that when the I p value of a water body to be discriminated in the rectangular area is not less than the average I p value of the extracted inland aquaculture ponds in this area, it is considered that this water body is an inland aquaculture pond. When the water body to be discriminated meets the above conditions, it means that it is an aquaculture pond.

[0103] Finally, there is a type of water body that cannot be effectively distinguished from inland aquaculture ponds by the above rules. The water body that is prone to be misclassified as an aquaculture pond is an artificial reservoir, because the artificial reservoir is also a regularly constructed water body. Figure 13 (a) is the water body under the satellite image, Figure 13 and the blue and red areas in (b) represent the artificial reservoir

[0104] and the adjacent inland aquaculture ponds respectively. The obvious difference between the two is the quantity and area size. Since the reservoir area is larger than that of general water bodies and due to the effect of water storage, the reservoirs are usually far apart. Therefore, this application uses the idea of discriminating by the maximum value of the water body area to extract aquaculture ponds. Statistically calculate and sort the areas of all water bodies in the rectangular area in ascending order to form a set, denoted as G, and we can get:

[0105] G = {S1, S2,..., S i ,...S m} (17)

[0106] where S iRepresents the connected domain area of the water body in the rectangular area aquaculture pond. Then, it is necessary to perform area discrimination on Sm. In statistics, if a set of data follows a normal distribution, under the 3-sigma principle, an outlier is a value in a set of measured values whose deviation from the mean exceeds 3 times the standard deviation. Here, it is assumed that the data in G follows a Gaussian distribution. Considering the huge difference in the connected domain area between a single aquaculture pond and an artificial reservoir and only discriminating the maximum value of a set of data, this application uses the ratio of the standard deviation of a set of data to the standard deviation of the set of data after removing the maximum value to measure such area differences, which can be expressed as:

[0107]

[0108] Among them, and are the means of {S1, S2,..., S i ,... S m} and {S1, S2,..., S i ,... S m-1} respectively. D ff is a measure of the degree of difference between a set of data before and after removing the maximum value, that is, the multiple of difference. According to the actual water body area difference, 5 times is taken here. When the difference between the above two sets of data exceeds 5, that is, the above relationship is not satisfied, it indicates that the rectangular area contains an artificial reservoir. Then, remove the water body with the largest connected domain area, and the water body result of the inland aquaculture pond extraction can be obtained.

[0109] Next, through specific example comparisons, the solutions and effects of the present invention will be further described.

[0110] To quantitatively evaluate the extraction performance of the aquaculture pond, the overlap ratio (Intersection Over Union, IOU)

[18] and the Kappa coefficient are used as accuracy indicators. IOU is the repetition rate of the extraction result and the ground truth map, which can be expressed as:

[0111]

[0112] Among them, TP is the number of correctly classified positive pixels, TN is the number of correctly classified negative pixels, FP is the number of misclassified positive pixels, and FN is the number of misclassified negative pixels. This evaluation index refers to the proportion of correctly predicted pixels in the total pixels, that is, the accuracy rate. When the value of IOU is larger, it indicates that the error between the extraction result of the algorithm and the ground truth map is smaller. The Kappa coefficient is a commonly used classification performance index obtained from the confusion matrix and can also be used to measure the classification accuracy, which can be expressed as:

[0113]

[0114] Among them, p0 represents the sum of the number of correctly classified samples in each category divided by the total number of samples, and can be expressed as:

[0115]

[0116] p e represents the sum of the products of the actual and predicted quantities corresponding to all categories, and then divided by the square of the total number of samples, and can be expressed as:

[0117]

[0118] Experimental result comparison and analysis

[0119] Experimental area 1 is located in Tianjin, and experimental area 2 is located in Zhongshan. The images for the experiments of this application are all local areas. Both of these areas have typical characteristics of inland aquaculture ponds, but there are differences in the main water body types. The river network in the Tianjin area is well-developed, while the water body 1 in the Zhongshan area is mainly lakes. In order to test the performance of the inland aquaculture pond extraction algorithm of this application, the inland aquaculture pond extraction algorithm based on the regularity of the water body edge contour (Water Contour RegularityBoundary basedInland Aquaculture Pond Extraction Algorithm, WRIP) is used as a comparison method.

[0120] As shown in Figure 14 for the Tianjin area, Figure 14 (a) is the RGB false color image of the Tianjin area. The water body types in this area are rich, and aquaculture ponds are densely distributed at the intersections of the main and branch rivers and around large artificial reservoirs; Figure 14 (b) is the ground truth map of the aquaculture ponds, where the green-covered area is the aquaculture pond distribution area; Figure 14 (c) and Figure 14 (d) are the aquaculture pond extraction results of the comparison algorithm WRIP and the method of this application respectively. Among them, the red, yellow, and blue rectangular frames respectively represent three different extraction result situations.

[0121] First, it can be seen from the blue rectangular frame area that for the relatively obvious large-scale aquaculture pond areas, both the method of this application and the WRIP algorithm can extract them. However, by comparing the local areas in the rectangular frame, it can be found that some rivers truncated due to pixel reasons or dam intervals are misclassified as aquaculture ponds by WRIP. The reason is that the area of the truncated river is small and its edge is relatively regular, resulting in incorrect classification of the area and zero curvature by WRIP. Secondly, it can be seen from the yellow rectangular frame area that the method of this application misses some aquaculture ponds. This is because the dams in this area are not obvious, resulting in multiple aquaculture ponds being connected together, so that they do not meet the spatial characteristics of aquaculture ponds and are removed. Finally, it can be seen from the red rectangular frame area that the method of this application misextracts some artificial reservoirs. The reason is that the water body in this area is adjacent to the river and is rectangular in shape, so that it meets the requirements for aquaculture pond extraction in terms of direction and regularity and is misclassified. It should be noted that although WRIP does not extract the reservoir, due to its classification using area and perimeter, the connected aquaculture ponds are removed because of their excessive area and perimeter, which indicates that the WRIP algorithm has poor discrimination between reservoirs and aquaculture ponds.

[0122] The Zhongshan area is as Figure 15 shown. Figure 15 (a) is the RGB false color image of the Zhongshan area. This area is rich in vegetation and contains a large natural lake. Most of the aquaculture ponds here are distributed on the edge of the lake due to low construction cost and convenient water diversion, etc.; Figure 15 (b) is the ground truth map of the aquaculture ponds, where the green-covered area is the distribution area of the aquaculture ponds; Figure 15 (c) and 15(d) are the aquaculture pond extraction results of the comparative algorithm WRIP and this application respectively. It can be seen from the extraction results that both methods can obtain good results for the main structure of the aquaculture ponds in this area, and there are also certain degrees of misclassification and missed detection.

[0123] In order to more clearly show the details of the extraction results of coastal aquaculture ponds by different methods, in Figure 15 (c) and 15(d), different extraction results are marked with rectangular frames of three different colors: red, yellow, and blue. By Figure 15(a) It can be seen that the area corresponding to the red rectangular frame is mainly dominated by buildings, which appears purple-white in the RGB false-color image. For the convenience of residents' water use, a relatively large number of small ponds have been built in this area, and the distance between the ponds is evenly spaced. Compared with the WRIP algorithm, the algorithm of the present application can remove such isolated small ponds by utilizing the dense distribution characteristics of the aquaculture pond water body in the local space. Therefore, in this area, the water body interference of the algorithm of the present application is less, and the algorithm performs better. It can be seen from the area of the yellow rectangular frame that the WRIP algorithm missed some aquaculture ponds because the areas of a few aquaculture ponds were not uniform, resulting in their failure to meet the area characteristics of the water body and being missed. Since the algorithm of the present application considers the local area of the aquaculture pond, the water body contained therein is screened again in the local area, which can reduce the situation of aquaculture ponds being missed. In the area shown by the blue rectangular frame, compared with Figure 15 (b) It can be seen that there is no aquaculture pond here. The WRIP and the algorithm of the present application wrongly extracted this water body because this area is the intersection of a lake and a small river, and the terrain is winding, resulting in many parts of the river being truncated into short strips and being difficult to remove.

[0124] Table 1 Comparison of the experimental performance of the algorithm of the present application and the comparative algorithm

[0125]

[0126] To intuitively compare the performance of the two methods, the present application uses two evaluation indicators, IOU and Kappa coefficient, to quantitatively analyze the extraction results, as shown in Table 1. In the Tianjin and Zhongshan regions, the IOU is about 80%, and the kappa is about 85%. Compared with the comparative algorithm, both the IOU and the Kappa coefficient have increased by more than 5%. It can be seen from the data in the table that the reason why the two algorithms perform better in the Tianjin region is that the scale of the inland aquaculture ponds in this region is relatively large and they are built more regularly, with obvious strip texture characteristics. The accuracy decreases in the Zhongshan region because the water body conditions in this region are complex, there are a large number of ponds and they are scattered, making it difficult to distinguish the aquaculture ponds with relatively small scales in terms of the connected domain area and shape characteristics.

[0127] In Landsat-8 optical remote sensing images, due to the complex types of inland water bodies, the low pixel resolution of the images, and the large number of small interfering water bodies, the extraction of inland aquaculture ponds is very challenging. In addition, the truncated rivers, reservoirs, ponds, and inland aquaculture ponds have different similarities in terms of spatial characteristics and geometric characteristics, etc., which increases the difficulty of extracting inland aquaculture ponds in Landsat-8 images.

[0128] To solve the above-mentioned problems, the present application proposes an inland aquaculture pond extraction algorithm based on angle constraint and double-neighborhood statistics. This algorithm extracts inland aquaculture ponds by combining the double-angle curvature feature with the direction and double-neighborhood characteristics of the spatial distribution of inland aquaculture ponds. The extraction results of the algorithm of the present application and the comparative algorithm for inland aquaculture ponds are compared through experiments, and the performance of the two algorithms in different regions and different water body types is analyzed. To verify the effectiveness of the algorithm of the present application, two evaluation indicators, namely the kappa coefficient and the IOU, are used to quantify the experimental results. The experiments show that, compared with the comparative algorithm, the present application has at least a 5% improvement in both evaluation indicators.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An extraction method for inland aquaculture ponds based on angle constraint and double-neighborhood statistics, characterized in that Including: Obtain Landsat-8 images, preprocess the Landsat-8 images, and extract water areas from the preprocessed images; Connect the broken rivers in the extracted water areas according to the similarity of the tangent angle directions in the local space between the broken rivers, so as to obtain river areas; Extract the aquaculture ponds near the river areas, including: Obtain the curvature of each pixel along the water body boundary, Construct a curvature feature based on the curvature of each pixel, and the curvature feature is used to describe the regularity of the water body contour, Construct an extraction feature according to the direction characteristics of the aquaculture ponds and the river direction, and the extraction feature is used to describe the spatial relationship between the aquaculture ponds and the rivers, Compare the curvature feature and the extraction feature extracted from the image with the preset curvature threshold and extraction threshold, and extract the area that meets the requirements as the extracted river-type aquaculture pond area; Extract the aquaculture ponds around lakes and reservoirs; Discriminate the aquaculture areas in the river areas, the aquaculture areas around lakes and reservoirs from three aspects: direction similarity, geometric shape similarity and area singular value, and finally obtain the extraction result of inland aquaculture ponds.

2. The method for extracting inland aquaculture ponds based on angle constraint and double-neighborhood statistics according to claim 1, wherein Preprocess the Landsat-8 images, including: performing radiometric calibration and atmospheric correction on the Landsat-8 images.

3. A method for extracting inland aquaculture ponds based on angle constraints and double-neighborhood statistics according to claim 1, characterized in that Connect the broken rivers in the extracted water areas according to the similarity of the tangent angle directions in the local space between the broken rivers, including: Obtain the endpoints of the rivers to be connected, select a local neighborhood within the range of the endpoints, and obtain another point on the river within the local neighborhood; Use the average slope of the local neighborhood space and the mean value of the tangent slopes at the endpoints to characterize the direction of the truncated river, so as to obtain the tangent angle representing the river flow direction; When the tangent angles of the river flow directions at the two endpoints of the river truncation satisfy the preset relationship, connect the two endpoints.

4. A method for extracting inland aquaculture ponds based on angle constraints and double-neighborhood statistics according to claim 1, wherein Connect the broken rivers in the extracted water areas according to the similarity of the tangent angle directions in the local space between the broken rivers, including: Make tangents at the endpoints of the two rivers to be connected, obtain the intersection point of the two tangents, and use the two river endpoints and the intersection point as endpoints to construct a triangular area to be connected; Connect the rivers according to the number of broken river segments in the area to be connected, including: When the number of river segments is not less than 1, determine the connection order according to the distance measurement of each river segment, When the number of river segments is 0, use the midpoint of the triangle median line as the transition point for river connection to connect the rivers.

5. A method for extracting inland aquaculture ponds based on angle constraint and double-neighborhood statistics according to claim 1, characterized in that, Extract the aquaculture ponds around lakes and reservoirs, including: Obtain the curvature of each pixel along the water body boundary; Construct a curvature feature based on the curvature of each pixel, and the curvature feature is used to describe the regularity of the water body contour; Judge that if a connected domain with a curvature feature meeting the preset requirements contains another connected domain with the same curvature feature meeting the preset requirements in its local space, then judge that the current connected domain is a lake aquaculture area.

Images