Film-covered tobacco field area estimation system based on deep learning and high-resolution remote sensing images

Abstract

The present invention relates to the technical field of image detection of film-coated tobacco fields, and in particular to a film-coated tobacco field area estimation system based on deep learning and high-resolution remote sensing images. The present invention obtains all tobacco field blocks and corresponding grayscale run matrices in an image acquisition module; analyzes the run texture information in an image analysis module to obtain the film texture feature parameters of each tobacco field block; and clusters all tobacco field blocks in an image processing module to obtain the film area through filtering. The present invention first divides the tobacco field remote sensing image into blocks, further uses the run information to analyze the color information and texture distribution of the tobacco field film, quantifies the film texture feature parameters of each tobacco field block, and then clusters the tobacco field blocks with similar film texture features for unified filtering, thereby reducing the damage to the image texture features caused by the film texture differences in different tobacco fields during the filtering process, improving the filtering effect, and thus improving the estimation accuracy of the film-coated tobacco field area.

Description

Technical Field

[0001] The present invention relates to the technical field of film-covered tobacco field image detection, and in particular to a film-covered tobacco field area estimation system based on deep learning and high-resolution remote sensing images. Background Art

[0002] In modern agricultural production, film covering technology is widely used in flue-cured tobacco planting. This technology helps to improve the soil environment, maintain soil temperature and moisture, inhibit weed growth, and increase tobacco yield and quality. In order to better manage planting resources and agricultural machinery operation efficiency, it is necessary to accurately estimate the area of ​​film-covered tobacco fields. At present, the estimation of the area of ​​film-covered tobacco fields mainly relies on remote sensing technology. High-resolution remote sensing images of tobacco fields are taken by drones, and deep learning methods are used to mark and identify the film-covered areas, so as to accurately estimate the area of ​​film-covered tobacco fields.

[0003] However, in the process of collecting remote sensing images of film-covered tobacco fields, noise may exist in the images of film-covered tobacco fields due to environmental factors or collection equipment, so filtering and noise reduction are required; but the existing technology mainly uses weighted estimation of pixel information in the local neighborhood and adjusts the pixel value to achieve the filtering purpose when filtering tobacco field images, while film-covered tobacco fields are artificially transformed agricultural land, and the texture characteristics of different tobacco field areas may vary greatly. The filtering method based on local pixel information may destroy the original texture characteristics, resulting in problems such as adhesion of the film texture or incomplete noise removal, which in turn leads to low estimation accuracy of the area of ​​film-covered tobacco fields. Summary of the invention

[0004] In order to solve the technical problem in the prior art that the filtering effect of tobacco field remote sensing images is poor and thus the estimation accuracy of the film-covered tobacco field area is low, the purpose of the present invention is to provide a film-covered tobacco field area estimation system based on deep learning and high-resolution remote sensing images. The technical scheme adopted is as follows:

[0005] The present invention proposes a film-covered tobacco field area estimation system based on deep learning and high-resolution remote sensing images, the system comprising:

[0006] Image acquisition module: used to obtain all tobacco field blocks in the tobacco field remote sensing image, and obtain the grayscale run matrix of each tobacco field block in a preset run direction;

[0007] Image analysis module: used for obtaining the film texture expression of each matrix element according to the element size of the matrix element in each grayscale run matrix and the color information of the pixel points in all runs; obtaining the film texture discreteness of each matrix element according to the distribution of all runs corresponding to each matrix element in each grayscale run matrix in the corresponding tobacco field block; obtaining the film texture feature parameters corresponding to the tobacco field block according to the film texture expression and the film texture discreteness of all matrix elements in each grayscale run matrix;

[0008] Image processing module: used to obtain all tobacco field block clusters in the tobacco field remote sensing image according to the film texture feature parameters and spatial distance of each tobacco field block; filter each tobacco field block cluster to obtain a filtered tobacco field remote sensing image; and obtain the film area in the tobacco field corresponding to the filtered tobacco field remote sensing image based on a deep learning algorithm.

[0009] Furthermore, the method for obtaining the coating texture expression degree includes:

[0010] In the tobacco field block corresponding to each grayscale run matrix, the coating color parameter of each run is obtained according to the brightness and saturation of all pixel points corresponding to each run, wherein the brightness is positively correlated with the coating color parameter, and the saturation is negatively correlated with the coating color parameter;

[0011] The coating color expression of each matrix element is determined according to the concentration of the coating color parameters of all runs corresponding to each matrix element; in each grayscale run matrix, the coating texture expression of each matrix element is obtained according to the element size of each matrix element and the coating color expression.

[0012] Furthermore, the method for obtaining the coating texture expression of each matrix element according to the element size of each matrix element and the coating color expression comprises:

[0013] In each of the grayscale run length matrices, according to the degree of deviation of the coating color expression of each matrix element relative to the coating color expression of all matrix elements in the same column, a column expression parameter of each matrix element is obtained;

[0014] In each of the grayscale run length matrices, a row expression parameter of each matrix element is obtained according to the deviation degree of the element size of each matrix element relative to the element sizes of all matrix elements in the same row;

[0015] The coating texture expression degree of each matrix element is obtained according to the column expression parameter and the row expression parameter of each matrix element; the column expression parameter and the row expression parameter are both positively correlated with the coating texture expression degree.

[0016] Furthermore, the method for obtaining the discreteness of the coating texture includes:

[0017] In each of the grayscale run matrices, obtain the run centroids of all runs corresponding to each matrix element and the run center of each run, and calculate the horizontal distance and vertical distance of the run center of each run relative to the run centroid;

[0018] According to the shape features of the tobacco field block corresponding to each grayscale run matrix and in combination with the preset run direction, the maximum reference length of each run is obtained; according to the length difference between the run length of each run and the maximum reference length, the coating direction coefficient of each run is obtained; the length difference is negatively correlated with the coating direction coefficient;

[0019] Obtaining a horizontal weight and a vertical weight according to the coating direction coefficient, wherein the sum of the horizontal weight and the vertical weight is 1; weighting the horizontal distance using the horizontal weight, and weighting the vertical distance using the vertical weight, and taking the weighted sum value as a discrete parameter of each stroke;

[0020] The discrete parameters of all runs in each matrix element are integrated to obtain the discreteness of the film texture of each matrix element.

[0021] Furthermore, the method for obtaining the maximum reference length includes:

[0022] In each of the tobacco field image blocks, the maximum distance between any two pixel points in a preset run direction is used as the maximum reference length of all runs in the corresponding tobacco field image block.

[0023] Furthermore, the method for acquiring the coating texture characteristic parameters includes:

[0024] Multiply the film texture expression of each matrix element by the film texture discreteness to obtain the film texture characteristic sub-parameters of each matrix element; average the film texture characteristic sub-parameters of all matrix elements in each grayscale run matrix to obtain the film texture characteristic parameters of each grayscale run matrix corresponding to the tobacco field block.

[0025] Furthermore, the method for obtaining the tobacco field block clusters includes:

[0026] According to the differences in the characteristic parameters of the film texture between different tobacco field blocks and the spatial distances, the metric distances between different tobacco field blocks are constructed; based on the metric distances, K-means clustering is performed on all the tobacco field blocks, and the optimal K value is obtained based on the elbow method to obtain clusters of all tobacco field blocks.

[0027] Furthermore, the method for acquiring the filtered tobacco field remote sensing image includes:

[0028] All the tobacco field blocks in each tobacco field block cluster are merged as a search window, and a neighborhood window is constructed with each pixel point as the center. The neighborhood window is the minimum inscribed square of the tobacco field block to which the corresponding pixel point belongs. A non-local mean filtering algorithm is used to filter each pixel point in all the search windows to obtain a filtered tobacco field remote sensing image.

[0029] Furthermore, the method for obtaining the coating area includes:

[0030] Constructing a training set of a preset semantic segmentation model, training the preset semantic segmentation model to obtain a trained preset semantic segmentation model; using the trained preset semantic segmentation model to obtain the film-covered area in the filtered tobacco field remote sensing image;

[0031] Based on the ground sampling distance parameter of the tobacco field remote sensing image, the film-covered area of ​​the film-covered region in the filtered tobacco field remote sensing image in the corresponding tobacco field is calculated.

[0032] Furthermore, the method for obtaining the tobacco field block includes:

[0033] Superpixel segmentation is performed on the tobacco field remote sensing image to obtain all tobacco field blocks.

[0034] The present invention has the following beneficial effects:

[0035] The present invention obtains all tobacco field blocks in a tobacco field remote sensing image in an image acquisition module, and obtains a grayscale run matrix of each tobacco field block in a preset run direction; in an image analysis module, according to the element size of the matrix elements in each grayscale run matrix and the color information of the corresponding pixel points in all runs, obtains the film texture expression of each matrix element; according to the distribution of all runs corresponding to each matrix element in each grayscale run matrix in the corresponding tobacco field block, obtains the film texture discreteness of each matrix element; according to the film texture expression and the film texture discreteness of all matrix elements in each grayscale run matrix, obtains the film texture feature parameters of the corresponding tobacco field block, and the closer the film texture feature parameters are, the more similar the corresponding film texture trends are; in an image processing module, according to the film texture feature parameters and spatial distance of each tobacco field block, obtains all tobacco field block clusters in the tobacco field remote sensing image; filters each tobacco field block cluster to obtain a filtered tobacco field remote sensing image; and obtains the film area in the tobacco field corresponding to the filtered tobacco field remote sensing image based on a deep learning algorithm. The present invention first divides the remote sensing image of the tobacco field into blocks, further uses the run-length information to analyze the color information and texture distribution of the tobacco field film, quantifies the film texture feature parameters of each tobacco field block, and then clusters the tobacco field blocks with similar film texture features for unified filtering, thereby reducing the damage to the image texture features caused by the film texture differences in different tobacco fields during the filtering process, improving the filtering effect, and further improving the estimation accuracy of the film-covered tobacco field area. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 A system block diagram of a film-covered tobacco field area estimation system based on deep learning and high-resolution remote sensing images provided by an embodiment of the present invention;

[0038] Figure 2 A flowchart of a method for obtaining a coating texture expression degree provided by an embodiment of the present invention;

[0039] Figure 3 A flowchart of a method for obtaining discreteness of film texture provided by an embodiment of the present invention;

[0040] Figure 4 An unfiltered tobacco field remote sensing image provided by one embodiment of the present invention;

[0041] Figure 5A filtered tobacco field remote sensing image provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the specific implementation method, structure, features and effects of a film-covered tobacco field area estimation system based on deep learning and high-resolution remote sensing images proposed by the present invention in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0043] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0044] The following is a detailed description of a specific scheme of a film-covered tobacco field area estimation system based on deep learning and high-resolution remote sensing images provided by the present invention in conjunction with the accompanying drawings.

[0045] See also Figure 1 , which shows a system block diagram of a film-covered tobacco field area estimation system based on deep learning and high-resolution remote sensing images provided by an embodiment of the present invention. The system includes: an image acquisition module 101, an image analysis module 102, and an image processing module 103.

[0046] Image acquisition module 101: used to acquire all tobacco field blocks in the tobacco field remote sensing image, and acquire the grayscale run length matrix of each tobacco field block in a preset run length direction.

[0047] In one embodiment of the present invention, a remote sensing image of a tobacco field of a tobacco field area to be estimated for film covering area is collected by using a drone equipped with a remote sensing sensor, wherein the tobacco field remote sensing image is an RGB image. Considering that film covering is affected by planting factors, different tobacco planting areas have different texture features presented in the tobacco field remote sensing image, which in turn affects the filtering effect, so the tobacco field remote sensing image is further segmented to obtain all tobacco field blocks, so as to facilitate the subsequent merging of tobacco field blocks with similar texture features for unified filtering, thereby obtaining high-quality filtered tobacco field remote sensing images to accurately estimate the film covering area.

[0048] It should be noted that care should be taken to avoid the impact of bad weather such as rain, fog, snow, etc. on the collection of remote sensing images of tobacco fields.

[0049] Preferably, in one embodiment of the present invention, a simple linear iterative clustering (SLIC) algorithm is specifically used to perform superpixel segmentation on the tobacco field remote sensing image, and each superpixel block is a tobacco field block. The SLIC algorithm is a well-known technology, and implementers may also use other superpixel segmentation algorithms, which will not be described in detail here.

[0050] In another embodiment of the present invention, the implementer may also, based on a deep learning algorithm, annotate the boundaries of tobacco fields for a large number of rich tobacco field remote sensing images acquired historically or downloaded from the Internet, and construct a segmentation training sample set to train a segmentation network model, which may be a U-Net model, etc.; then, the currently acquired tobacco field remote sensing image is input into the trained segmentation network model, and all tobacco field blocks in the tobacco field remote sensing image can be obtained; this is a well-known technology and will not be elaborated here.

[0051] In order to facilitate the subsequent analysis of the texture information of the tobacco field remote sensing image and evaluate whether it has the film-like texture features, the embodiment of the present invention further obtains the grayscale run matrix of each tobacco field block in a preset run direction. In one embodiment of the present invention, the RGB image of the tobacco field remote sensing image is first converted into a grayscale image, and the tobacco field blocks segmented in the RGB image are projected into the grayscale image to obtain the tobacco field blocks in the grayscale image; then the grayscale level is divided into 16 grayscale intervals, that is, the grayscale range is 0-256, and each 16 grayscale values ​​are a grayscale interval, and the preset run direction is set to 0°, so as to obtain all the runs of each tobacco field block in the grayscale image, and then construct the grayscale run matrix.

[0052] In the grayscale run matrix, the run refers to the continuous sequence corresponding to the pixel points in the same grayscale interval in the preset run direction; different matrix elements on the same row represent the statistical values ​​of the pixel point sequence in the same grayscale interval but with different run lengths, and different matrix elements on the same column represent the statistical values ​​of the pixel point sequence in different grayscale intervals but with the same run length. The statistical value is the total number of such runs.

[0053] It should be noted that the grayscale run matrix is ​​a well-known technology, and implementers can divide the grayscale intervals into other numbers and can also set other run directions by themselves, which will not be elaborated here.

[0054] Image analysis module 102: used to obtain the film texture expression of each matrix element according to the element size of the matrix element in each grayscale run matrix and the color information of the corresponding pixel points in all runs; obtain the film texture discreteness of each matrix element according to the distribution of all runs corresponding to each matrix element in each grayscale run matrix in the corresponding tobacco field block; obtain the film texture feature parameters of the corresponding tobacco field block according to the film texture expression and film texture discreteness of all matrix elements in each grayscale run matrix.

[0055] Taking into account whether the film covering direction is consistent with the preset run direction, the regular texture of the film covering area makes the corresponding runs similar, that is, they belong to the same type of runs, which makes the corresponding statistical values ​​in the grayscale run matrix, that is, the matrix elements, larger; but because the film covering usually indirectly causes similar regular texture characteristics in the soil, the matrix elements corresponding to the runs in the soil area in the grayscale run matrix are also larger, making it difficult to evaluate whether the runs corresponding to the matrix elements are film-covered runs; and considering that the ground film covered in the tobacco field area is usually white or black, it has a distinctive color characteristic relative to the surrounding soil or vegetation, that is, film-covered runs usually have special color characteristics.

[0056] Therefore, the embodiment of the present invention obtains the coating texture expression of each matrix element according to the element size of the matrix element in each grayscale run matrix and the color information of the corresponding pixel points in all runs; the coating texture expression of the matrix element is comprehensively evaluated from two perspectives: color and run statistics, i.e., matrix elements, to quantify the coating texture characteristics of each tobacco field block, so as to subsequently combine all matrix elements to evaluate the similarity of coating textures of different tobacco field blocks, so that each pixel point in the tobacco field remote sensing image can refer to the local neighborhood pixel information with similar texture characteristics during the filtering process, thereby reducing the damage to the image texture characteristics and improving the filtering effect.

[0057] Preferably, in one embodiment of the present invention, the method for obtaining the expression degree of the coating texture includes:

[0058] See also Figure 2 , which shows a flow chart of a method for obtaining a coating texture expression degree provided by an embodiment of the present invention, specifically comprising:

[0059] Step S201, in each grayscale run matrix corresponding to the tobacco field block, according to the brightness and saturation of all pixels corresponding to each run, obtain the coating color parameter of each run, the brightness is positively correlated with the coating color parameter, and the saturation is negatively correlated with the coating color parameter.

[0060] Taking into account that the surface of the ground film covered in the tobacco field is smooth and will reflect light, resulting in a higher brightness in the remote sensing image of the tobacco field; and considering that the ground film is usually black or white with a low saturation; in one embodiment of the present invention, the RGB image is first converted into the HSV space to evaluate the brightness and saturation of the image; and because the higher the brightness and the lower the saturation, the more it conforms to the texture characteristics of the film, the brightness is positively normalized and the saturation is negatively normalized.

[0061] As an example, the calculation formula for the coating color parameter is: Among them, Q m For each grayscale run matrix, the coating color parameter of the mth run in the tobacco field block corresponds to the mth run; i is the sequence number of the pixel in the mth run; l is the total number of pixels in the mth run; norm() is the standard normalization function, specifically using linear normalization; V i is the brightness of the i-th pixel in the m-th run; exp() is an exponential function with the natural constant e as the base; S i is the saturation of the i-th pixel in the m-th run; exp(-S i ) indicates S i Negative correlation normalization.

[0062] In another embodiment, the implementer may also replace the saturation with the gray value. Since the color of the ground film is white or black, the closer its gray value is to 0 or 255, the greater the difference with the median of the gray range, indicating that the color characteristics of the film are more obvious, and |G i -127| replaces exp(-S i ), where G i is the gray value of the i-th pixel in the m-th run, thereby obtaining the coating color parameter of each run.

[0063] It should be noted that HSV conversion is a well-known technology and will not be described in detail here. The implementer may also adopt other normalization methods, or may obtain the film color parameters of each run by evaluating the difference between the saturation or hue of each pixel and the corresponding color of the film according to actual needs, such as when the film is of other colors, and will not be described in detail here.

[0064] Step S202, determine the coating color expression of each matrix element according to the concentration of the coating color parameters of all runs corresponding to each matrix element; in each grayscale run matrix, obtain the coating texture expression of each matrix element according to the element size and coating color expression of each matrix element.

[0065] Since each matrix element represents a type of run, after obtaining the coating color parameters of each run, the coating color expression of each matrix element can be evaluated. In one embodiment of the present invention, the mean of all runs corresponding to each matrix element is used as the coating color expression of each matrix element, and the mode can also be used instead of the mean to represent the concentration. The coating color expression of each matrix element is evaluated, which will not be elaborated here.

[0066] In the grayscale run matrix, considering the regular texture of the film-covered area and the soil area, their corresponding runs are similar and numerous, which usually makes the corresponding matrix elements have larger values ​​and relatively deviate from other matrix elements on the same row; and considering the film color expression of the corresponding matrix elements of the film-covered runs, it will be relatively large and deviate from other matrix elements on the same column.

[0067] Therefore, in a preferred embodiment of the present invention, according to the element size and coating color expression of each matrix element, the method for obtaining the coating texture expression of each matrix element includes:

[0068] In each grayscale run-length matrix, the column expression parameter of each matrix element is obtained according to the degree of deviation of the film color expression of each matrix element relative to the film color expression of all matrix elements in the same column; in each grayscale run-length matrix, the row expression parameter of each matrix element is obtained according to the degree of deviation of the element size of each matrix element relative to the element size of all matrix elements in the same row; the film texture expression of each matrix element is obtained according to the column expression parameter and row expression parameter of each matrix element; both the column expression parameter and the row expression parameter are positively correlated with the film texture expression.

[0069] As an example, the calculation formula for the expression of the coating texture includes: Among them, F k is the coating texture expression of the kth matrix element in each grayscale run matrix; norm() is the standard normalization function; Q k is the coating color expression of the kth matrix element in each grayscale run matrix; is the mean value of the coating color expression of all matrix elements in the column to which the kth matrix element belongs in each grayscale run matrix; X k is the element value of the kth matrix element in each grayscale run matrix; is the mean value of all matrix elements in the column to which the kth matrix element belongs in each grayscale run matrix; is the list expression parameter of the kth matrix element in each grayscale run matrix; Express the parameters for the row of the k-th matrix element in each grayscale run matrix.

[0070] In the above formula, the deviation degree of any value is evaluated by the difference between the value and the average value, so as to obtain the column expression parameter and row expression parameter of each matrix element respectively. The larger the difference is, the larger the normalized value is, which means that the matrix element is more likely to correspond to the coating texture. Then the two are multiplied and combined to obtain the coating texture expression degree.

[0071] In other examples, implementers may also replace the average with the maximum value, perform negative correlation mapping on the difference, evaluate the degree of deviation, and obtain the column expression parameters and row expression parameters of each matrix element respectively; implementers may also use basic mathematical operations such as addition or weighted summation or related mapping methods to combine the two to obtain the coating texture expression, which will not be elaborated here.

[0072] At this point, the film texture expression of each matrix element in the grayscale run-length matrix corresponding to each tobacco field block can be obtained, so as to subsequently combine the film texture discreteness to quantify the film texture feature parameters of each tobacco field block.

[0073] Taking into account that the runs corresponding to the film texture should be distributed regularly and discretely in the image, the embodiment of the present invention obtains the discreteness of the film texture of each matrix element according to the distribution of all runs corresponding to each matrix element in each grayscale run matrix in the corresponding tobacco field block; the discreteness of the film texture is further combined with the run direction and distribution to quantify the film texture characteristics.

[0074] Preferably, in one embodiment of the present invention, the method for obtaining the discreteness of the coating texture includes:

[0075] See also Figure 3 , which shows a flow chart of a method for obtaining the discreteness of a coating texture provided by an embodiment of the present invention, specifically comprising:

[0076] Step S301, in each grayscale run matrix, obtain the run centroid of all runs corresponding to each matrix element and the run center of each run, and calculate the horizontal distance and vertical distance of the run center of each run relative to the run centroid.

[0077] In order to analyze the discrete distribution of the itinerary, we first obtain relevant information parameters to prepare for subsequent analysis.

[0078] As an example, first obtain the centroid of each run corresponding to each matrix element in all runs, then average the corresponding coordinates of all centroids to obtain the run centroid of all runs corresponding to each matrix element; synchronously obtain the run center of each run, where the run centroid and the run center both correspond to a pixel point in the tobacco field tile, and the two may overlap; further obtain the horizontal distance and vertical distance of the run center of each run relative to the run centroid, and both the horizontal distance and the vertical distance reflect the distribution of all runs corresponding to each matrix element.

[0079] In other examples, implementers may also use the center of gravity instead of the center of mass, or directly average the coordinates corresponding to the center of each run instead of the run center of mass; these are all existing technologies well known to those skilled in the art and will not be described in detail here.

[0080] It should be noted that when calculating the horizontal distance and the vertical distance, a two-dimensional coordinate system is constructed with the center of the remote sensing image of the entire tobacco field as the origin to determine the coordinates of the pixel point; the horizontal distance and the vertical distance are based on the coordinate system and on the preset stroke direction. When the preset stroke direction is 0°, it is consistent with the horizontal axis direction of the coordinate.

[0081] Step S302, according to the shape characteristics of the tobacco field block corresponding to each grayscale run matrix and combined with the preset run direction, the maximum reference length of each run is obtained; according to the length difference between the run length of each run and the maximum reference length, the coating direction coefficient of each run is obtained; the length difference is negatively correlated with the coating direction coefficient.

[0082] Taking into account whether the coating direction is consistent with the preset run direction, there is a certain influence on the discrete distribution law of the evaluation run; and taking into account that the coated area in the tobacco field is arranged in a regular strip shape, for the run representing the coating texture, the larger the run length, the greater the possibility that the run direction is closer to the coating texture direction; and it is difficult to determine whether it represents the coating texture based on the run length of a single run; therefore, an embodiment of the present invention first obtains the maximum reference length of the run in each tobacco field block, and the maximum reference length reflects the maximum possible length of the run corresponding to the coated area in the tobacco field block. When the run length is closer to the maximum reference length, it means that the run direction is closer to the coating texture direction, and thus the coating direction coefficient can be obtained.

[0083] In a preferred embodiment of the present invention, a method for obtaining the maximum reference length includes: in each tobacco field block, taking the maximum distance between any two pixel points in a preset run direction as the maximum reference length of all runs in the corresponding tobacco field block, wherein the maximum reference length is always greater than or equal to the run length of all runs in the tobacco field block.

[0084] As an example, the absolute value of the difference between the stroke length and the maximum reference length is negatively correlated and mapped. Specifically, the absolute value of the difference is used as the x in the exponential function exp(-x) with the natural constant e as the base. Other negative correlation mapping methods can also be used. The smaller the absolute value of the difference, the closer the stroke direction of the stroke is to the coating texture direction, and the larger the coating direction coefficient.

[0085] In another example, the implementer may also evaluate the coating direction coefficient by the ratio of the run length to the maximum reference length. The closer the ratio is to 1, the closer the run direction of the run is to the coating texture direction, and the larger the coating direction coefficient is.

[0086] Step S303, obtaining horizontal weight and vertical weight according to the coating direction coefficient, the sum of the horizontal weight and the vertical weight is 1; using the horizontal weight to weight the horizontal distance, and the vertical weight to weight the vertical distance, and taking the weighted sum value as the discrete parameter of each stroke.

[0087] Considering that the larger the coating direction coefficient is, the greater the possibility that each run corresponding to each matrix element is a run in a whole coating, the horizontal coordinates of the run center of each run and the run centroid corresponding to all runs are relatively close, and the degree of distribution discreteness is mainly reflected by the vertical distance between the run center and the run centroid; and when the coating direction coefficient is smaller, the run can only represent part of the coating area, and it is necessary to comprehensively analyze and evaluate its distribution discreteness by horizontal distance and vertical distance.

[0088] As an example, the discrete parameter is calculated as: In the formula, The k-th matrix element in each grayscale run matrix corresponds to the discrete parameter of the m-th run among all runs; The kth matrix element in each grayscale run matrix corresponds to the coating direction coefficient of the mth run among all runs, that is, the vertical weight; The k-th matrix element in each grayscale run matrix corresponds to the horizontal weight of the m-th run among all runs; The k-th matrix element in each grayscale run matrix corresponds to the horizontal distance between the run center of the m-th run and the run centroid among all runs; The kth matrix element in each grayscale run matrix corresponds to the vertical distance between the run center of the mth run and the run centroid among all runs.

[0089] Step S304, synthesizing the discrete parameters of all runs in each matrix element to obtain the discreteness of the film texture of each matrix element.

[0090] As an example, the discrete parameters of all runs in each matrix element are averaged to obtain the discreteness of the film texture of each matrix element; in other examples, the mode can also be used instead of the mean, which will not be described in detail here.

[0091] At this point, the discreteness of the film texture of each matrix element in the grayscale run matrix corresponding to each tobacco field block is obtained; the embodiment of the present invention further obtains the film texture feature parameters of the corresponding tobacco field block based on the film texture expression and the film texture discreteness of all matrix elements in each grayscale run matrix.

[0092] Preferably, in one embodiment of the present invention, the method for acquiring the coating texture characteristic parameters includes:

[0093] The film texture expression of each matrix element is multiplied by the film texture discreteness to obtain the film texture characteristic sub-parameters of each matrix element; the film texture characteristic sub-parameters of all matrix elements in each grayscale run matrix are averaged to obtain the film texture characteristic parameters of the tobacco field block corresponding to each grayscale run matrix.

[0094] As an example, the calculation formula of the coating texture feature parameter is: Where, T s is the film texture feature parameter of the tobacco field block corresponding to the sth grayscale run matrix; k is the sequence number of the matrix element in the sth grayscale run matrix; n is the total number of matrix elements in the sth grayscale run matrix; F k,s M is the coating texture expression of the kth matrix element in the sth grayscale run matrix; k,s is the film texture discreteness of the kth matrix element in the sth grayscale run matrix; F k,s ×M k,s It is the film texture feature sub-parameter of the kth matrix element in the sth grayscale run-length matrix.

[0095] In other embodiments of the present invention, implementers may also use basic mathematical operations such as addition or weighted summation or related mapping methods to combine the film texture expression and the film texture discreteness to obtain the film texture characteristic sub-parameters; or they may use the mode of all film texture characteristic sub-parameters instead of the mean, which will not be repeated here.

[0096] At this point, the film texture feature parameters of each tobacco field block are obtained, so that tobacco field blocks with similar film textures can be filtered together in the image processing module 103, reducing the damage of neighborhood pixel information to image texture features and improving filtering effect.

[0097] Image processing module 103: used to obtain all tobacco field block clusters in the tobacco field remote sensing image according to the film texture feature parameters and spatial distance of each tobacco field block; filter each tobacco field block cluster to obtain a filtered tobacco field remote sensing image; and obtain the film area in the tobacco field corresponding to the filtered tobacco field remote sensing image based on a deep learning algorithm.

[0098] Considering that the film texture feature parameters of each tobacco field block are quantified in the image analysis module 102, the closer the film texture feature parameters are, the more similar the corresponding film texture trends are; however, in order to merge tobacco field blocks with similar texture features for unified filtering, the spatial distance of the tobacco field blocks needs to be further considered.

[0099] Preferably, in one embodiment of the present invention, the method for obtaining tobacco field block clusters includes:

[0100] According to the differences in the characteristic parameters of the film texture and the spatial distance between different tobacco field blocks, the metric distance between different tobacco field blocks was constructed; K-means clustering was performed on all tobacco field blocks based on the metric distance, and the optimal K value was obtained based on the elbow method to obtain the clustering of all tobacco field blocks.

[0101] As an example, for any two tobacco field tiles, the distance between the two tobacco field tiles is measured according to the Euclidean distance formula. The calculation formula for measuring the distance is specifically expressed as: Among them, D s,s+1 is the metric distance between the tobacco field blocks corresponding to the sth grayscale run matrix and the s+1th grayscale run matrix; T s is the film texture feature parameter of the tobacco field block corresponding to the sth grayscale run matrix; T s+1 is the film texture feature parameter of the tobacco field block corresponding to the s+1th grayscale run matrix; B s,s+1 It is the spatial distance between the centers of the tobacco field blocks corresponding to the sth grayscale run matrix and the s+1th grayscale run matrix; in other examples, the difference in the feature parameters of the film texture and the spatial distance can also be weighted and averaged to obtain the metric distance.

[0102] It should be noted that the K-means algorithm and the elbow method are already existing technologies well known to those skilled in the art, and implementers may also use other clustering methods, which will not be described in detail here.

[0103] After obtaining all the tobacco field tile clusters, they can be further filtered and denoised to obtain high-quality filtered tobacco field remote sensing images.

[0104] Preferably, in one embodiment of the present invention, considering that non-local means filtering can better preserve edge and texture information of an image, a method for acquiring a filtered tobacco field remote sensing image includes:

[0105] All tobacco field blocks in each tobacco field block cluster are merged as the search window, and a neighborhood window is constructed with each pixel as the center. The neighborhood window is the minimum inscribed square of the tobacco field block to which the corresponding pixel belongs. The non-local mean filtering algorithm is used to filter each pixel in all search windows to obtain the filtered tobacco field remote sensing image.

[0106] It should be noted that the non-local mean filtering algorithm is an existing technology well known to those skilled in the art and will not be described in detail here.

[0107] See also Figure 4 and Figure 5 , which respectively show an unfiltered tobacco field remote sensing image and a filtered tobacco field remote sensing image provided by an embodiment of the present invention; Figure 4 There is a certain amount of salt and pepper noise in the middle film area and the corresponding texture of some film areas is blurred and lost; Figure 5 After filtering, the remote sensing images of the tobacco fields have clearer textures and can accurately estimate the film-covered area.

[0108] After obtaining high-quality filtered tobacco field remote sensing images, the film covering area in the tobacco field corresponding to the filtered tobacco field remote sensing images can be obtained based on the deep learning algorithm.

[0109] Preferably, in one embodiment of the present invention, the method for obtaining the coating area includes:

[0110] Construct a training set of a preset semantic segmentation model, train the preset semantic segmentation model, and obtain a trained preset semantic segmentation model; use the trained preset semantic segmentation model to obtain the film-covered area in the filtered tobacco field remote sensing image; based on the ground sampling distance parameter of the tobacco field remote sensing image, calculate the film-covered area in the corresponding tobacco field in the filtered tobacco field remote sensing image.

[0111] As an example, the film-covered areas in a large number of rich filtered tobacco field remote sensing images acquired historically are manually annotated, and the annotated filtered tobacco field remote sensing images are divided into training set and validation set according to a ratio of 7:3. The preset semantic segmentation model is set to the DeeplabV3+ model for training, and the loss function is the cross entropy function. The network parameters are updated using the back propagation algorithm, and the training is performed until the loss function converges. The training is completed and the validation set is input for verification to obtain the trained preset semantic segmentation model; then the currently acquired tobacco field remote sensing image is input into the trained preset semantic segmentation model to obtain all the film-covered areas in the currently acquired tobacco field remote sensing image; then based on the ground sampling distance parameter of the tobacco field remote sensing image, the film-covered area of ​​the film-covered area in the filtered tobacco field remote sensing image in the corresponding tobacco field is calculated.

[0112] It should be noted that the training and application of semantic segmentation models and the calculation of the coverage area based on ground sampling distance parameters are already well-known technologies. Implementers can also train and apply other semantic segmentation models, which will not be elaborated here.

[0113] In summary, the embodiment of the present invention obtains all tobacco field blocks in the tobacco field remote sensing image in the image acquisition module, and obtains the grayscale run matrix of each tobacco field block in a preset run direction; in the image analysis module, according to the element size of the matrix elements in each grayscale run matrix and the color information of the corresponding pixel points in all runs, the film texture expression of each matrix element is obtained; according to the distribution of all runs corresponding to each matrix element in each grayscale run matrix in the corresponding tobacco field block, the film texture discreteness of each matrix element is obtained; according to the film texture expression and film texture discreteness of all matrix elements in each grayscale run matrix, the film texture feature parameters of the corresponding tobacco field block are obtained; in the image processing module, according to the film texture feature parameters and spatial distance of each tobacco field block, all tobacco field block clusters in the tobacco field remote sensing image are obtained; each tobacco field block cluster is filtered to obtain a filtered tobacco field remote sensing image; based on the deep learning algorithm, the film area in the tobacco field corresponding to the filtered tobacco field remote sensing image is obtained. The present invention first divides the remote sensing image of the tobacco field into blocks, further uses the run-length information to analyze the color information and texture distribution of the tobacco field film, quantifies the film texture feature parameters of each tobacco field block, and then clusters the tobacco field blocks with similar film texture features for unified filtering, thereby reducing the damage to the image texture features, improving the filtering effect, and thus improving the estimation accuracy of the film-covered tobacco field area.

[0114] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0115] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

Claims

1. A film-covered tobacco field area estimation system based on deep learning and high-resolution remote sensing images, characterized in that: The system comprises: Image acquisition module: used to obtain all tobacco field blocks in the tobacco field remote sensing image, and obtain the grayscale run matrix of each tobacco field block in a preset run direction; Image analysis module: used for obtaining the film texture expression of each matrix element according to the element size of the matrix element in each grayscale run matrix and the color information of the pixel points in all runs; obtaining the film texture discreteness of each matrix element according to the distribution of all runs corresponding to each matrix element in each grayscale run matrix in the corresponding tobacco field block; obtaining the film texture feature parameters corresponding to the tobacco field block according to the film texture expression and the film texture discreteness of all matrix elements in each grayscale run matrix; Image processing module: used for obtaining all tobacco field block clusters in the tobacco field remote sensing image according to the film texture feature parameters and spatial distance of each tobacco field block; filtering each tobacco field block cluster to obtain a filtered tobacco field remote sensing image; obtaining the film covering area in the tobacco field corresponding to the filtered tobacco field remote sensing image based on a deep learning algorithm; The method for obtaining the coating texture expression degree includes: In the tobacco field block corresponding to each grayscale run matrix, the coating color parameter of each run is obtained according to the brightness and saturation of all pixel points corresponding to each run, wherein the brightness is positively correlated with the coating color parameter, and the saturation is negatively correlated with the coating color parameter; Determine the coating color expression of each matrix element according to the concentration of the coating color parameters of all runs corresponding to each matrix element; obtain the coating texture expression of each matrix element according to the element size of each matrix element and the coating color expression in each grayscale run matrix; The method for obtaining the coating texture expression of each matrix element according to the element size of each matrix element and the coating color expression comprises: In each of the grayscale run length matrices, according to the degree of deviation of the coating color expression of each matrix element relative to the coating color expression of all matrix elements in the same column, a column expression parameter of each matrix element is obtained; In each of the grayscale run length matrices, a row expression parameter of each matrix element is obtained according to the deviation degree of the element size of each matrix element relative to the element sizes of all matrix elements in the same row; According to the column expression parameter and the row expression parameter of each matrix element, the coating texture expression degree of each matrix element is obtained; the column expression parameter and the row expression parameter are both positively correlated with the coating texture expression degree; The method for obtaining the discreteness of the coating texture includes: In each of the grayscale run matrices, obtain the run centroids of all runs corresponding to each matrix element and the run center of each run, and calculate the horizontal distance and vertical distance of the run center of each run relative to the run centroid; According to the shape features of the tobacco field block corresponding to each grayscale run matrix and in combination with the preset run direction, the maximum reference length of each run is obtained; according to the length difference between the run length of each run and the maximum reference length, the coating direction coefficient of each run is obtained; the length difference is negatively correlated with the coating direction coefficient; Obtaining a horizontal weight and a vertical weight according to the coating direction coefficient, wherein the sum of the horizontal weight and the vertical weight is 1; weighting the horizontal distance using the horizontal weight, and weighting the vertical distance using the vertical weight, and taking the weighted sum value as a discrete parameter of each stroke; The discrete parameters of all runs in each matrix element are integrated to obtain the discreteness of the film texture of each matrix element.

2. The system for estimating the area of ​​film-covered tobacco fields based on deep learning and high-resolution remote sensing images according to claim 1 is characterized in that: The method for obtaining the maximum reference length includes: In each of the tobacco field image blocks, the maximum distance between any two pixel points in a preset run direction is used as the maximum reference length of all runs in the corresponding tobacco field image block.

3. The system for estimating the area of ​​film-covered tobacco fields based on deep learning and high-resolution remote sensing images according to claim 1 is characterized in that: The method for obtaining the coating texture characteristic parameters includes: Multiply the film texture expression of each matrix element by the film texture discreteness to obtain the film texture characteristic sub-parameters of each matrix element; average the film texture characteristic sub-parameters of all matrix elements in each grayscale run matrix to obtain the film texture characteristic parameters of each grayscale run matrix corresponding to the tobacco field block.

4. The system for estimating the area of ​​film-covered tobacco fields based on deep learning and high-resolution remote sensing images according to claim 1 is characterized in that: The method for obtaining the tobacco field block clusters comprises: According to the differences in the characteristic parameters of the film texture between different tobacco field blocks and the spatial distances, the metric distances between different tobacco field blocks are constructed; based on the metric distances, K-means clustering is performed on all the tobacco field blocks, and the optimal K value is obtained based on the elbow method to obtain clusters of all tobacco field blocks.

5. The system for estimating the area of ​​film-covered tobacco fields based on deep learning and high-resolution remote sensing images according to claim 1, characterized in that: The method for acquiring the filtered tobacco field remote sensing image comprises: All the tobacco field blocks in each tobacco field block cluster are merged as a search window, and a neighborhood window is constructed with each pixel point as the center. The neighborhood window is the minimum inscribed square of the tobacco field block to which the corresponding pixel point belongs. A non-local mean filtering algorithm is used to filter each pixel point in all the search windows to obtain a filtered tobacco field remote sensing image.

6. The system for estimating the area of ​​film-covered tobacco fields based on deep learning and high-resolution remote sensing images according to claim 1 is characterized in that: The method for obtaining the film covering area includes: Constructing a training set of a preset semantic segmentation model, training the preset semantic segmentation model to obtain a trained preset semantic segmentation model; using the trained preset semantic segmentation model to obtain the film-covered area in the filtered tobacco field remote sensing image; Based on the ground sampling distance parameter of the tobacco field remote sensing image, the film-covered area of ​​the film-covered region in the filtered tobacco field remote sensing image in the corresponding tobacco field is calculated.

7. The system for estimating the area of ​​film-covered tobacco fields based on deep learning and high-resolution remote sensing images according to claim 1, characterized in that: The method for obtaining the tobacco field block includes: Superpixel segmentation is performed on the tobacco field remote sensing image to obtain all tobacco field blocks.

Images