CN113223042B - Intelligent acquisition method and equipment for remote sensing image deep learning sample - Google Patents
Intelligent acquisition method and equipment for remote sensing image deep learning sample Download PDFInfo
- Publication number
- CN113223042B CN113223042B CN202110544661.0A CN202110544661A CN113223042B CN 113223042 B CN113223042 B CN 113223042B CN 202110544661 A CN202110544661 A CN 202110544661A CN 113223042 B CN113223042 B CN 113223042B
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- sensing image
- sample
- data
- pixel
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/181—Segmentation; Edge detection involving edge growing; involving edge linking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30181—Earth observation
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to the field of remote sensing image acquisition, and discloses a method and equipment for intelligently acquiring a remote sensing image deep learning sample, which comprises the steps of acquiring remote sensing image data, and judging whether the remote sensing image data has remote sensing image target marking data of a corresponding time phase; generating remote sensing image target marking data by using an image segmentation energy model; judging whether the space coordinate systems of the remote sensing image data and the remote sensing image target labeling data are consistent or not; cutting the remote sensing image data and the remote sensing image target marking data, regularizing the pixel values of the remote sensing image target marking data according to a classification system, and marking category codes; screening an initial sample data set of the remote sensing image, and carrying out regular naming; and storing the standard sample set into a sample database according to the sample storage structure. The method can perform standard processing aiming at various types of remote sensing image data and labeled data, and generates the remote sensing image deep learning sample set which is classified by rules and easy to uniformly manage.
Description
Technical Field
The invention relates to the technical field of remote sensing image acquisition and processing, in particular to an intelligent acquisition method and equipment for a remote sensing image deep learning sample.
Background
In recent years, with the increasing number of satellites in China, remote sensing technology is rapidly developed, remote sensing images are increasingly abundant in types, the updating speed is accelerated, the image data volume is also increased explosively, and the remote sensing technology enters a big data era. However, the traditional remote sensing interpretation and information extraction are generally low in efficiency and cannot meet the requirement of big data application, so how to rapidly and intelligently identify and extract information in remote sensing big data also becomes a new development direction in the technical field of remote sensing.
Meanwhile, artificial intelligence research taking deep learning as a core is rapidly developed, various image processing deep learning algorithms appear in the field of computer vision, and various image recognition and information extraction algorithms are increasingly applied to processing and extracting remote sensing images, so that great progress is made in the aspects of image classification, target recognition, change discovery and the like of big data of the remote sensing images, and the development of intelligent information extraction of the remote sensing images is promoted.
In deep learning, a large number of training samples are needed to learn data characteristics, and optimal model parameters are obtained to ensure the accuracy of a result model. However, the current remote sensing data has various scales and varieties, interpretation labeling information of the remote sensing data is also various, the requirements of the sample are different along with the requirements of the model and the extraction category, various standard processing needs to be carried out on the data in the manufacturing process of the sample, and the sample acquisition and manufacturing process is complicated.
Disclosure of Invention
The invention provides an intelligent acquisition method and equipment for a remote sensing image deep learning sample, and aims to solve the problems in the prior art.
In a first aspect, the invention provides an intelligent acquisition method for a remote sensing image deep learning sample, which comprises the following steps:
s1), obtaining remote sensing image data, judging whether the remote sensing image data has remote sensing image target marking data of a corresponding time phase, and turning to the step S2 if the remote sensing image data does not have the remote sensing image target marking data of the corresponding time phase); if the remote sensing image target labeling data of the corresponding time phase exists, turning to the step S3);
s2), semi-automatic interactive labeling acquisition is carried out, a foreground seed line is drawn in a part of target area selected from remote sensing image data through manual interaction in an operation window, and the foreground seed line is used as a foreground example; establishing an image segmentation energy model, and performing parameter estimation on the image segmentation energy model by using a foreground example; deducing the remote sensing image data by using an image segmentation energy model to generate remote sensing image target labeling data;
s3) judging whether the space coordinate system of the remote sensing image data is consistent with the space coordinate system of the remote sensing image target labeling data, if so, turning to the step S4); if not, performing geographical reference modification or spatial projection modification on the remote sensing image target annotation data according to the spatial coordinate system of the remote sensing image data to ensure that the sample data is consistent in space, and turning to the step S4);
s4), setting a sample size and a sample data format, calculating a sample acquisition step length, respectively cutting the remote sensing image data and the remote sensing image target marking data according to the sample size and the sample acquisition step length, judging whether the remote sensing image target marking data is in a vector form, if so, rasterizing the remote sensing image target marking data, and regularizing the pixel value of the rasterized remote sensing image target marking data according to a classification system and marking a class code; if not, directly regularizing the pixel values of the remote sensing image target labeling data according to a classification system and labeling class codes; obtaining a group of remote sensing image sample pair data, repeating the steps until the remote sensing image data and the remote sensing image target marking data are cut, and obtaining an initial sample data set of the remote sensing image; proceeding to step S5);
s5) screening the remote sensing image initial sample data set, removing sample data with low sample pixel ratio in the remote sensing image initial sample data set, and obtaining a screened sample data set;
s6) carrying out regularized naming on the screened sample data set to obtain a standard sample set;
s7) storing the standard sample set into the sample database according to the sample storage structure.
Further, in step S2), an image segmentation energy model is established, and the foreground example is used to perform parameter estimation on the image segmentation energy model; the method for deducing the remote sensing image data by utilizing the image segmentation energy model to generate the remote sensing image target annotation data comprises the following steps:
s21) assuming that the classification label of the pixels of the entire image of the remote sensing image data is L ═ L1,l2,…,li,…,lpP represents the total number of pixel points, liLabel, l, representing the ith pixeli0 denotes background,/i1 denotes the target; assuming that the whole image of the remote sensing image data is segmented into L, establishing an image segmentation energy model E (L) ═ aR (L) + B (L) according to the whole image segmentation of the remote sensing image data, wherein R (L) is an area item, B (L) is a boundary item, E (L) represents image segmentation energy, and a represents an important factor between the area item and the boundary item;
s22) establishing a Gaussian mixture model aiming at the region item, wherein the Gaussian mixture model comprises k Gaussian models, RGB three-channel vectors of each pixel of the remote sensing image data are obtained, the probability of each pixel in the remote sensing image data being classified into a target is respectively calculated through the RGB three-channel vectors and the k Gaussian models, the maximum probability value obtained through calculation in the k Gaussian models is taken as the classification result of the pixel, and the region item is obtained according to the classification result of each pixel;
s23) establishing a pixel difference model aiming at the boundary item, wherein the pixel difference model uses a BGR three-channel vector to measure the similarity of two pixels, the Euclidean distance is adopted to calculate the pixel difference value between two pixel points, if the pixel difference value between a certain pixel point and an adjacent pixel point is larger, the edge between the certain pixel point and the adjacent pixel point is used as a cut edge, and the boundary item is obtained according to the cut edge;
s24) calculating an image segmentation energy function according to the region item obtained in the step S23) and the boundary item obtained in the step S24), and repeating the steps S23) to S24) until the image segmentation energy function converges to the minimum value, so as to obtain an optimal image segmentation energy model;
s25) the remote sensing image is segmented by utilizing the optimal image segmentation energy model, and target labeling data of the remote sensing image are obtained.
Further, in step S22), the probability that each pixel in the remote sensing image data is classified into the target is calculated through the RGB three-channel vector and using k gaussian models, the maximum value of the probability calculated in the k gaussian models is taken, the maximum value of the probability is used as the classification result of the pixel, the region item is obtained according to the classification result of each pixel, and the probability that the ith pixel is classified into the target is further included
Wherein x represents BGR three-channel vector of ith pixel point, pijThe ratio of the number Ni of the pixel point samples input to the jth Gaussian model to the total number N of the pixel point samples is expressed,
and 0 is less than or equal to pij≤1；gj(x；μj,Σj) A probability model representing the jth gaussian model,
μj、Σjrespectively representing mean value and covariance matrix obtained by BGR three-channel vector calculation of all pixel point samples input to the jth Gaussian model, and mean value muj＝(μB,μG,μR)。
Further, in step S4), the method includes clipping the remote-sensing image data and the remote-sensing image target labeling data according to the sample size and the sample collection step size, respectively, and includes calculating, for the remote-sensing image data and the remote-sensing image target labeling data, the row and column numbers corresponding to four fixed points of a single sample on the whole remote-sensing image by using the sample size and the sample collection step size, and calculating corresponding spatial coordinates according to the corresponding row and column numbers, where the spatial coordinates include an X-direction geospatial coordinate Xgeo and a Y-direction geospatial coordinate Ygeo, the X-direction geospatial coordinate Xgeo [0] + Xpixel Geo [1] + Ypixel [ Geo [2], the Y-direction geospatial coordinate Ygeo [3] + Xpixel [4] + Ypixel [5], where Geo [0] and Geo [3] are the upper-corner geographic coordinate and the Y-coordinate, and Geo [1] Geo [5] are respectively related to the east-west-geographic coordinate of the image, and the X-direction And north-south image resolution, Geo 2 and Geo 4 are north-south image resolution and east-west image resolution respectively associated with y-direction geospatial coordinates, Xpixel and Ypixel respectively represent line number and column number of the image; obtaining a sample space range according to the space coordinates; and cutting the remote sensing image data and the remote sensing image target marking data according to the sample space range.
Further, in step S4), regularizing according to a classification system, where the classification system includes 2 different levels of land types, and the 2 different levels of land types include a first level land type and a second level land type, where the first level land type includes a water body, vegetation, a road, a building area, an open land, glaciers, and perennial snow; the secondary land types corresponding to the water body comprise oceans, rivers, lakes, ponds and paddy fields; secondary land species corresponding to vegetation include dry land, garden land and grassland; the secondary land types corresponding to the roads comprise asphalt roads, cement roads, earth and stone roads and railways; the secondary land types corresponding to the building areas comprise independent houses, rural residential points, town building groups and other structures; the secondary land types corresponding to the bare land include fallow land, sand land, miners' bare land and mixed bare land; the second class of land corresponding to glaciers and perennial snow includes glaciers and perennial snow.
Further, in step S6), performing a regulated naming on the filtered sample data set, where the regulated naming includes a remote sensing image name, a sample target type, and a sample serial number.
Further, in step S7), the sample storage structure includes sample data information and sample metadata information, where the sample data information includes a sample data type, a sample data description, and a sample data constraint mode; the sample metadata information includes a sample metadata type, a sample metadata description, and a sample metadata constraint mode.
On the other hand, the invention provides intelligent acquisition equipment for a remote sensing image deep learning sample, which comprises the following components: the intelligent acquisition method comprises the steps of a memory, a processor and an intelligent acquisition program of the remote sensing image deep learning sample, wherein the intelligent acquisition program of the remote sensing image deep learning sample is stored in the memory and can be operated on the processor, and the intelligent acquisition method of the remote sensing image deep learning sample is realized when the intelligent acquisition program of the remote sensing image deep learning sample is executed by the processor.
The invention has the beneficial effects that: the invention can utilize the image segmentation energy model to deduce the remote sensing image data and generate the remote sensing image target marking data, in addition, the invention carries out geographical reference modification or spatial projection modification on the remote sensing image target marking data according to the spatial coordinate system of the remote sensing image data, thus ensuring the consistency of sample data space.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments are briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.
Fig. 1 is a schematic flow chart of an intelligent acquisition method for a remote sensing image deep learning sample according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings. It should be understood that the detailed description and specific examples, while indicating the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention. It is noted that the terms "comprises" and "comprising," and any variations thereof, in the description and claims of the present invention and the above-described drawings are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of elements is not necessarily limited to those elements, but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.
The first embodiment, the first aspect, the invention provides an intelligent acquisition method for a remote sensing image deep learning sample, which comprises the following steps:
s1), obtaining remote sensing image data, judging whether the remote sensing image data has remote sensing image target marking data of a corresponding time phase, and turning to the step S2 if the remote sensing image data does not have the remote sensing image target marking data of the corresponding time phase); if the remote sensing image target labeling data of the corresponding time phase exists, turning to the step S3);
s2), semi-automatic interactive labeling acquisition is carried out, a foreground seed line is drawn in a part of target area selected from remote sensing image data through manual interaction in an operation window, and the foreground seed line is used as a foreground example; establishing an image segmentation energy model, and performing parameter estimation on the image segmentation energy model by using a foreground example; deducing the remote sensing image data by using an image segmentation energy model to generate remote sensing image target labeling data;
in step S2), an image segmentation energy model is established, and parameter estimation is performed on the image segmentation energy model by using the foreground example; the method for deducing the remote sensing image data by utilizing the image segmentation energy model to generate the remote sensing image target annotation data comprises the following steps:
s21) assuming that the classification label of the pixels of the entire image of the remote sensing image data is L ═ L1,l2,…,li,…,lpP represents the total number of pixel points, liLabel, l, representing the ith pixeli0 denotes background,/i1 denotes the target; assuming that the whole image of the remote sensing image data is segmented into L, establishing an image segmentation energy model E (L) ═ aR (L) + B (L) according to the whole image segmentation of the remote sensing image data, wherein R (L) is an area item, B (L) is a boundary item, E (L) represents image segmentation energy, and a represents an important factor between the area item and the boundary item;
s22) establishing a Gaussian mixture model aiming at the region item, wherein the Gaussian mixture model comprises k Gaussian models, RGB three-channel vectors of each pixel of the remote sensing image data are obtained, the probability of each pixel in the remote sensing image data being classified into a target is respectively calculated through the RGB three-channel vectors and the k Gaussian models, the maximum probability value obtained through calculation in the k Gaussian models is taken as the classification result of the pixel, and the region item is obtained according to the classification result of each pixel;
in step S22), calculating probability of each pixel point in the remote sensing image data being classified into the target by using k gaussian models through RGB three-channel vectors, taking the maximum probability value calculated in the k gaussian models, taking the maximum probability value as the classification result of the pixel point, obtaining a region item according to the classification result of each pixel point, and further including the probability of the ith pixel point being classified into the target
Wherein x represents BGR three-channel vector of ith pixel point, pijThe ratio of the number Ni of the pixel point samples input to the jth Gaussian model to the total number N of the pixel point samples is expressed,
and 0 is less than or equal to pij≤1；
μj、ΣjRespectively representing mean value and covariance matrix obtained by BGR three-channel vector calculation of all pixel point samples input to the jth Gaussian model, and mean value muj＝(μB,μG,μR)。
In the image segmentation energy model, a Gaussian mixture model is adopted for calculating the regional items, in the embodiment, 5 Gaussian models are respectively arranged corresponding to the target and the background, when the probability that a certain pixel point belongs to the foreground is calculated, a plurality of Gaussian models are adopted for calculation, and the highest probability is selected as a final attribution result.
S23) establishing a pixel difference model aiming at the boundary item, wherein the pixel difference model uses BGR three-channel vector to measure the similarity of two pixels, the Euclidean distance is adopted to calculate the pixel difference value between two pixel points, if the pixel difference value between a certain pixel point and an adjacent pixel point is larger, the edge between the certain pixel point and the adjacent pixel point is used as a cut edge, and the boundary item is obtained according to the cut edge.
S24) calculating an image segmentation energy function according to the region item obtained in the step S23) and the boundary item obtained in the step S24), and repeating the steps S23) to S24) until the image segmentation energy function converges to the minimum value, so as to obtain an optimal image segmentation energy model;
the boundary term is calculated by measuring the similarity of two pixels through a BGR three channel, and the difference between the two pixels is calculated by adopting the Euclidean distance. The larger the pixel difference between a certain pixel point and an adjacent pixel point is, the smaller the overall energy is when the edge between two pixels is a cut edge. Meanwhile, through iterative minimization, parameters of a Gaussian mixture model for modeling the target and the background are optimized in each iterative process, and the image segmentation effect is improved.
In the embodiment, an image segmentation energy model is established, and parameter estimation is performed on the image segmentation energy model by using a target area example (namely a foreground example); and segmenting the remote sensing image data by utilizing the established image segmentation energy model. The image segmentation energy model utilizes texture (color) information and boundary (contrast) information in an image, selects some target areas and background areas through a small amount of user interaction operation, and segments and extracts the image. The process of the image segmentation energy model to distinguish the target from the background can be obtained by minimizing the energy function by minimizing the image segmentation. The energy of the cut edge occurring at the boundary of the object and the background is minimal. The region item reflects the overall characteristics of the pixel sample set, and the boundary item reflects the difference between two pixel points.
S25) the remote sensing image is segmented by utilizing the optimal image segmentation energy model, and target labeling data of the remote sensing image are obtained.
S3) judging whether the space coordinate system of the remote sensing image data is consistent with the space coordinate system of the remote sensing image target labeling data, if so, turning to the step S4); if not, performing geographical reference modification or spatial projection modification on the remote sensing image target annotation data according to the spatial coordinate system of the remote sensing image data to ensure that the sample data is consistent in space, and turning to the step S4);
s4), setting a sample size and a sample data format, calculating a sample acquisition step length, respectively cutting the remote sensing image data and the remote sensing image target marking data according to the sample size and the sample acquisition step length, judging whether the remote sensing image target marking data is in a vector form, if so, rasterizing the remote sensing image target marking data, and regularizing the pixel value of the rasterized remote sensing image target marking data according to a classification system and marking a class code; if not, directly regularizing the pixel values of the remote sensing image target labeling data according to a classification system and labeling class codes; obtaining a group of remote sensing image sample pair data, repeating the steps until the remote sensing image data and the remote sensing image target marking data are cut, and obtaining an initial sample data set of the remote sensing image; proceeding to step S5);
in step S4), the remote-sensing image data and the remote-sensing image target labeling data are respectively clipped according to the sample size and the sample collection step size, the method includes calculating the corresponding row and column numbers of four fixed points of a single sample on the whole remote-sensing image according to the sample size and the sample collection step size for the remote-sensing image data and the remote-sensing image target labeling data, and calculating the corresponding spatial coordinates according to the corresponding row and column numbers, where the spatial coordinates include an X-direction geospatial coordinate Xgeo and a Y-direction geospatial coordinate Ygeo, the X-direction geospatial coordinate Xgeo [0] + Xpixel [ Geo [1] + Ypixel [ Geo [2], the Y-direction geospatial coordinate Ygeo [3] + Xpixel [ Geo [4] + Ypixel [5], where Geo [0] and Geo [3] are the left upper-corner X coordinate of the image and the Y-geographic coordinate Geo [1] and the east-west-spatial resolution of the image in the X-direction and the X-direction, and the north-west-geographic resolution of the image are respectively related to the X-direction The image resolution, Geo 2 and Geo 4 are south-north direction image resolution and east-west direction image resolution related to y direction geographic space coordinate, Xpixel and Ypixel represent line number and column number of image; obtaining a sample space range according to the space coordinates; and cutting the remote sensing image data and the remote sensing image target marking data according to the sample space range.
When the remote sensing image data and the remote sensing image target marking data are cut, standard cutting needs to be carried out according to the size of a sample and the sample step length, the sample step length is the difference value of the overlapping distance between the size of a sample block and the sample block, the number of remote sensing image deep learning samples can be increased due to the overlapping of the sample blocks, the richness of the samples is improved, and the precision of model training is guaranteed.
In step S4), regularizing according to a classification system, wherein the classification system includes 2 different levels of land types, and the 2 different levels of land types include a first level land type and a second level land type, and the first level land type includes a water body, vegetation, roads, a building area, bare land, glaciers and perennial snow; the secondary land types corresponding to the water body comprise oceans, rivers, lakes, ponds and paddy fields; secondary land species corresponding to vegetation include dry land, garden land and grassland; the secondary land types corresponding to the roads comprise asphalt roads, cement roads, earth and stone roads and railways; the secondary land types corresponding to the building areas comprise independent houses, rural residential points, town building groups and other structures; the secondary land types corresponding to the bare land include fallow land, sand land, miners' bare land and mixed bare land; the second class of land corresponding to glaciers and perennial snow includes glaciers and perennial snow.
S5) screening the remote sensing image initial sample data set, removing sample data with low sample pixel ratio in the remote sensing image initial sample data set, and obtaining a screened sample data set;
s6) carrying out regularized naming on the screened sample data set to obtain a standard sample set; the regularized naming comprises a remote sensing image name, a sample target type and a sample serial number.
S7) storing the standard sample set into the sample database according to the sample storage structure. The sample storage structure comprises sample data information and sample metadata information, wherein the sample data information comprises a sample data type, a sample data description and a sample data constraint mode; the sample metadata information includes a sample metadata type, a sample metadata description, and a sample metadata constraint mode. In this embodiment, the sample picture may be directly stored, and if the standard sample set needs to be stored in the sample database, the attribute information of the sample set is added according to the sample storage rule of the sample database, and the sample set is stored in the sample database.
In this embodiment, the generated sample data set may directly store a picture file and save a sample data pair; and if the sample set needs to be stored in the sample database, adding the attribute information of the sample set according to the sample storage rule of the sample database, and storing the sample set in the sample database. The sample database stores attribute information such as image time phase, image resolution, sample type and the like of the remote sensing image deep learning sample together, so that operations such as inspection, retrieval, updating and the like can be conveniently performed on a large number of deep learning samples of various types, and the remote sensing image deep learning sample can be better managed.
On the other hand, the invention provides intelligent acquisition equipment for a remote sensing image deep learning sample, which comprises the following components: the intelligent acquisition method comprises the steps of a memory, a processor and an intelligent acquisition program of the remote sensing image deep learning sample, wherein the intelligent acquisition program of the remote sensing image deep learning sample is stored in the memory and can be operated on the processor, and the intelligent acquisition method of the remote sensing image deep learning sample is realized when the intelligent acquisition program of the remote sensing image deep learning sample is executed by the processor.
By adopting the technical scheme disclosed by the invention, the following beneficial effects are obtained:
the invention can utilize the image segmentation energy model to deduce the remote sensing image data and generate the remote sensing image target marking data, in addition, the invention carries out geographical reference modification or spatial projection modification on the remote sensing image target marking data according to the spatial coordinate system of the remote sensing image data, thus ensuring the consistency of sample data space.
The foregoing is only a preferred embodiment of the present invention, and it should be noted that it will be apparent to those skilled in the art that various modifications and improvements can be made without departing from the principle of the present invention, and such modifications and improvements should also be considered within the scope of the present invention.
Claims (7)
1. An intelligent acquisition method for a remote sensing image deep learning sample is characterized by comprising the following steps:
s1), obtaining remote sensing image data, judging whether the remote sensing image data has remote sensing image target marking data of a corresponding time phase, and turning to the step S2 if the remote sensing image data does not have the remote sensing image target marking data of the corresponding time phase); if the remote sensing image target labeling data of the corresponding time phase exists, turning to the step S3);
s2), semi-automatic interactive labeling acquisition is carried out, a foreground seed line is drawn in a part of target areas selected from the remote sensing image data through manual interaction in an operation window, and the foreground seed line is used as a foreground example; establishing an image segmentation energy model, and performing parameter estimation on the image segmentation energy model by using the foreground example; deducing the remote sensing image data by using the image segmentation energy model to generate remote sensing image target labeling data;
s3) judging whether the space coordinate system of the remote sensing image data is consistent with the space coordinate system of the remote sensing image target labeling data, if so, turning to the step S4); if not, performing geographical reference modification or spatial projection modification on the remote sensing image target annotation data according to the spatial coordinate system of the remote sensing image data to ensure that the sample data is consistent in space, and turning to the step S4);
s4), setting a sample size and a sample data format, calculating a sample acquisition step length, respectively cutting the remote sensing image data and the remote sensing image target marking data according to the sample size and the sample acquisition step length, judging whether the remote sensing image target marking data is in a vector form, if so, rasterizing the remote sensing image target marking data, and regularizing the pixel value of the rasterized remote sensing image target marking data according to a classification system and marking a class code; if not, directly regularizing the pixel value of the remote sensing image target labeling data according to a classification system and labeling a class code; obtaining a group of remote sensing image sample pair data, repeating the step until the remote sensing image data and the remote sensing image target marking data are cut, and obtaining an initial sample data set of the remote sensing image; proceeding to step S5);
s5) screening the remote sensing image initial sample data set, removing sample data with low sample pixel ratio in the remote sensing image initial sample data set, and obtaining a screened sample data set;
s6) carrying out regularized naming on the screened sample data set to obtain a standard sample set;
s7) storing the standard sample set into a sample database according to a sample storage structure;
in step S2), an image segmentation energy model is established, and parameter estimation is performed on the image segmentation energy model by using the foreground example; deducing the remote sensing image data by using the image segmentation energy model to generate remote sensing image target labeling data, comprising the following steps of:
s21) assuming that the classification label of the pixels of the entire image of the remote sensing image data is L ═ L1，l2，...，li，...，lpP represents the total number of pixel points, liLabel, l, representing the ith pixeli0 denotes background,/i1 denotes the target; assuming that the whole image of the remote sensing image data is segmented into L, establishing an image segmentation energy model E (L) ═ aR (L) + B (L) according to the whole image segmentation of the remote sensing image data, wherein R (L) is an area item, B (L) is a boundary item, E (L) represents image segmentation energy, and a represents an important factor between the area item and the boundary item;
s22) establishing a Gaussian mixture model aiming at the region item, wherein the Gaussian mixture model comprises k Gaussian models, RGB three-channel vectors of each pixel of the remote sensing image data are obtained, the probability of each pixel in the remote sensing image data being classified into a target is respectively calculated through the RGB three-channel vectors and the k Gaussian models, the maximum probability value obtained through calculation in the k Gaussian models is taken as the classification result of the pixel, and the region item is obtained according to the classification result of each pixel;
s23) establishing a pixel difference model aiming at the boundary item, wherein the pixel difference model measures the similarity of two pixels by using a BGR three-channel vector, the Euclidean distance is adopted to calculate the pixel difference value between two pixel points, if the pixel difference value between a certain pixel point and an adjacent pixel point is larger, the edge between the certain pixel point and the adjacent pixel point is used as a cut edge, and the boundary item is obtained according to the cut edge;
s24) calculating an image segmentation energy function according to the region item obtained in the step S23) and the boundary item obtained in the step S24), and repeating the steps S23) to S24) until the image segmentation energy function converges to the minimum value, so as to obtain an optimal image segmentation energy model;
s25) segmenting the remote sensing image by utilizing the optimal image segmentation energy model to obtain target annotation data of the remote sensing image.
2. The intelligent acquisition method for the remote sensing image deep learning sample according to claim 1, characterized in that in step S22), the probability of classifying each pixel in the remote sensing image data into a target is respectively calculated through the RGB three-channel vector and the k gaussian models, the maximum probability value calculated in the k gaussian models is taken, the maximum probability value is taken as the classification result of the pixel, a region item is obtained according to the classification result of each pixel, and the method further comprises the step of classifying the ith pixel into the target
Wherein x represents BGR three-channel vector of ith pixel point, pijThe ratio of the number Ni of the pixel point samples input to the jth Gaussian model to the total number N of the pixel point samples is expressed,
and 0 is less than or equal to pij≤1；gj(x；μj，∑j) A probability model representing the jth gaussian model,
μj、∑jand respectively representing a mean value and a covariance matrix obtained by BGR three-channel vector calculation of all pixel point samples input to the jth Gaussian model.
3. The method according to claim 2, wherein in step S4), the remote-sensing image data and the remote-sensing image target labeling data are respectively clipped according to the sample size and the sample collection step length, and the method comprises calculating, for the remote-sensing image data and the remote-sensing image target labeling data, the row and column numbers corresponding to four fixed points of a single sample on the whole remote-sensing image by using the sample size and the sample collection step length, and calculating the corresponding spatial coordinates according to the corresponding row and column numbers, wherein the spatial coordinates include x-direction geospatial coordinates Xgeo and y-direction geospatial coordinates Ygeo, the x-direction geospatial coordinates Xgeo [0] + Xpixel _ Geo [1] + Ypixel _ Geo [2], the y-direction geospatial coordinates Ygeo [4] + ygixel [5], wherein, Geo 0 and Geo 3 are X geographic coordinate and Y geographic coordinate of the upper left corner of the image respectively, Geo 1 and Geo 5 are east-west direction image resolution and south-north direction image resolution related to X direction geographic space coordinate respectively, Geo 2 and Geo 4 are south-north direction image resolution and east-west direction image resolution related to Y direction geographic space coordinate respectively, and Xpixel and Ypixel represent row number and column number of the image respectively; obtaining a sample space range according to the space coordinates; and cutting the remote sensing image data and the remote sensing image target marking data according to the sample space range.
4. The intelligent remote sensing image deep learning sample collection method according to claim 3, wherein in step S4), the classification system is regularized according to a classification system, the classification system comprises 2 different levels of land types, the 2 different levels of land types comprise a primary land type and a secondary land type, and the primary land type comprises water, vegetation, roads, construction areas, bare land, glaciers and perennial snow; the secondary land types corresponding to the water body comprise oceans, rivers, lakes, ponds and paddy fields; secondary land species corresponding to vegetation include dry land, garden land and grassland; the secondary land types corresponding to the roads comprise asphalt roads, cement roads, earth and stone roads and railways; the secondary land types corresponding to the building areas comprise independent houses, rural residential points, town building groups and other structures; the secondary land types corresponding to the bare land comprise fallow land, sand land, miner bare land and mixed bare land; the second grade land corresponding to glaciers and perennial snow cover includes glaciers and perennial snow cover.
5. The intelligent acquisition method for the remote sensing image deep learning samples according to claim 1, wherein in step S6), the filtered sample data set is subjected to regular naming, and the regular naming comprises a remote sensing image name, a sample target type and a sample serial number.
6. The intelligent acquisition method for the remote sensing image deep learning sample according to claim 1, wherein in step S7), the sample storage structure comprises sample data information and sample metadata information, and the sample data information comprises a sample data type, a sample data description and a sample data constraint mode; the sample metadata information includes a sample metadata type, a sample metadata description, and a sample metadata constraint mode.
7. An intelligent remote sensing image deep learning sample acquisition device, which is suitable for the intelligent remote sensing image deep learning sample acquisition method according to any one of claims 1 to 6, and is characterized by comprising the following components: the intelligent acquisition program for the remote sensing image deep learning samples is stored in the memory and can be operated on the processor, and when being executed by the processor, the intelligent acquisition program for the remote sensing image deep learning samples realizes the steps of the intelligent acquisition method for the remote sensing image deep learning samples as claimed in any one of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110544661.0A CN113223042B (en) | 2021-05-19 | 2021-05-19 | Intelligent acquisition method and equipment for remote sensing image deep learning sample |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110544661.0A CN113223042B (en) | 2021-05-19 | 2021-05-19 | Intelligent acquisition method and equipment for remote sensing image deep learning sample |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113223042A CN113223042A (en) | 2021-08-06 |
| CN113223042B true CN113223042B (en) | 2021-11-05 |
Family
ID=77092938
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110544661.0A Active CN113223042B (en) | 2021-05-19 | 2021-05-19 | Intelligent acquisition method and equipment for remote sensing image deep learning sample |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113223042B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113705640B (en) * | 2021-08-16 | 2024-03-01 | 中国电子科技集团公司第二十八研究所 | Method for quickly constructing airplane detection data set based on remote sensing image |
| CN113989678B (en) * | 2021-11-18 | 2022-06-10 | 生态环境部卫星环境应用中心 | Method and device for constructing ecosystem classification sample library |
| CN114170242A (en) * | 2021-11-22 | 2022-03-11 | 北京小米移动软件有限公司 | Sample image generation method, image prediction method, device, equipment and medium |
| CN114663790B (en) * | 2022-05-24 | 2022-08-26 | 济宁德信测绘有限公司 | Intelligent remote sensing mapping method and system |
| CN117349462B (en) * | 2023-12-06 | 2024-03-12 | 自然资源陕西省卫星应用技术中心 | Remote sensing intelligent interpretation sample data set generation method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105550639A (en) * | 2015-12-07 | 2016-05-04 | 国家测绘地理信息局卫星测绘应用中心 | Automatic extraction method for Earth observation laser height measurement satellite elevation control points and data processing method |
| CN112329751A (en) * | 2021-01-06 | 2021-02-05 | 北京道达天际科技有限公司 | Deep learning-based multi-scale remote sensing image target identification system and method |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9230168B2 (en) * | 2013-07-31 | 2016-01-05 | Digitalglobe, Inc. | Automatic generation of built-up layers from high resolution satellite image data |
| US9495618B1 (en) * | 2015-08-06 | 2016-11-15 | Digitalglobe, Inc. | Object detection with textural to spectral domain adaptation |
| CN106909644B (en) * | 2017-02-21 | 2019-03-26 | 中国科学院电子学研究所 | A kind of multistage tissue and indexing means towards mass remote sensing image |
| CN108363951B (en) * | 2018-01-11 | 2022-02-22 | 东南大学 | Automatic acquisition method of deep learning sample library corresponding to remote sensing image land type identification |
| CN111144487B (en) * | 2019-12-27 | 2023-09-26 | 二十一世纪空间技术应用股份有限公司 | Method for establishing and updating remote sensing image sample library |
| CN111582356B (en) * | 2020-04-30 | 2022-03-29 | 中国人民解放军海军航空大学 | Target information unified representation and association method |
| CN112102338A (en) * | 2020-09-17 | 2020-12-18 | 西安泽塔云科技股份有限公司 | Method and device for acquiring image sample data based on deep learning |
-
2021
- 2021-05-19 CN CN202110544661.0A patent/CN113223042B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105550639A (en) * | 2015-12-07 | 2016-05-04 | 国家测绘地理信息局卫星测绘应用中心 | Automatic extraction method for Earth observation laser height measurement satellite elevation control points and data processing method |
| CN112329751A (en) * | 2021-01-06 | 2021-02-05 | 北京道达天际科技有限公司 | Deep learning-based multi-scale remote sensing image target identification system and method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113223042A (en) | 2021-08-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113223042B (en) | Intelligent acquisition method and equipment for remote sensing image deep learning sample | |
| Shi et al. | Road detection from remote sensing images by generative adversarial networks | |
| Zhang et al. | A multilevel point-cluster-based discriminative feature for ALS point cloud classification | |
| Guan et al. | Integration of orthoimagery and lidar data for object-based urban thematic mapping using random forests | |
| CN103049763B (en) | Context-constraint-based target identification method | |
| CN107067405B (en) | Remote sensing image segmentation method based on scale optimization | |
| CN103268358B (en) | Multi-source control point image database builds and update method | |
| CN113449594A (en) | Multilayer network combined remote sensing image ground semantic segmentation and area calculation method | |
| CN106846322B (en) | The SAR image segmentation method learnt based on curve wave filter and convolutional coding structure | |
| CN111028255A (en) | Farmland area pre-screening method and device based on prior information and deep learning | |
| CN112347970A (en) | Remote sensing image ground object identification method based on graph convolution neural network | |
| CN113033520A (en) | Tree nematode disease wood identification method and system based on deep learning | |
| CN111738113A (en) | Road extraction method of high-resolution remote sensing image based on double-attention machine system and semantic constraint | |
| Guo et al. | Urban impervious surface extraction based on multi-features and random forest | |
| CN112329559A (en) | Method for detecting homestead target based on deep convolutional neural network | |
| CN113505842A (en) | Automatic extraction method suitable for large-scale regional remote sensing image urban building | |
| CN107563366A (en) | A kind of localization method and device, electronic equipment | |
| CN112241676A (en) | Method for automatically identifying terrain sundries | |
| Liu et al. | Spatial-temporal hidden Markov model for land cover classification using multitemporal satellite images | |
| CN109657728B (en) | Sample production method and model training method | |
| CN107392926B (en) | Remote sensing image feature selection method based on early-stage land thematic map | |
| CN113628180A (en) | Semantic segmentation network-based remote sensing building detection method and system | |
| CN116503750A (en) | Large-range remote sensing image rural block type residential area extraction method and system integrating target detection and visual attention mechanisms | |
| CN115082778B (en) | Multi-branch learning-based homestead identification method and system | |
| Tao et al. | An approach for identifying historic village using deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |