CN111144487B - Method for establishing and updating remote sensing image sample library - Google Patents
Method for establishing and updating remote sensing image sample library Download PDFInfo
- Publication number
- CN111144487B CN111144487B CN201911372783.5A CN201911372783A CN111144487B CN 111144487 B CN111144487 B CN 111144487B CN 201911372783 A CN201911372783 A CN 201911372783A CN 111144487 B CN111144487 B CN 111144487B
- Authority
- CN
- China
- Prior art keywords
- sample
- remote sensing
- sensing image
- data
- sample data
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
Abstract
The application discloses a method for establishing and updating a remote sensing image sample library, which comprises the following steps: when an operation instruction for establishing a remote sensing image sample library is received, establishing the remote sensing image sample library according to a first preset rule; when an instruction for extracting from a remote sensing image sample database is received, judging whether sample data meeting a first preset condition exists in a remote book database according to the extraction requirement; under the condition that sample data meeting a first preset condition does not exist in a sample library, receiving a label of a first object in the remote sensing image data, and extracting a second object in the remote sensing image data by using an iterative training and preset method according to the label so as to update the sample library; under the condition that sample data meeting a first preset condition exists in a sample library, extracting the sample data in the sample library, performing model training, remote sensing image data prediction and optimization processing, extracting incremental samples according to a second preset rule, and updating the incremental samples into the sample library.
Description
Technical Field
The application relates to the technical field of artificial intelligence, in particular to a method for establishing and updating a remote sensing image sample library.
Background
In recent years, with rapid development of Deep learning technology, particularly, there is a great progress in the fields of remote sensing image information extraction and the like, and the premise of Deep Neural NetWorks (DNNs) corresponding to Deep learning is that the network is sufficiently trained, and a large number of samples are required as training data when training the Deep Neural NetWorks.
The sample labeling is an important part of training data in deep learning, and when labeling the remote sensing image, various ground object samples in the remote sensing image are required to be labeled and integrated to form a sample library. The sample library stores labeling vector data and basic period image files of various ground features, and corresponding data can be called from the sample library according to sample types when the sample library is used so as to train DNN.
However, the labeling of the samples in the related technology generally depends on manual labeling, a large amount of manpower and material resources are consumed, the labeling efficiency is low, and the effect is not ideal; the traditional sample organization management mode does not use a spatial database, and is inconvenient for sample data management and use.
Disclosure of Invention
In view of the above, the application provides a method for establishing and updating a remote sensing image sample library, which aims to solve the problems of low labeling efficiency and unsatisfactory effect of manually labeling samples in remote sensing image data in the related technology.
In order to achieve the above object, according to one aspect of the present application, there is provided a method for creating and updating a remote sensing image sample library, including:
(1) When an operation instruction for establishing a remote sensing image sample library is received, establishing the remote sensing image sample library according to a first preset rule;
(2) When an instruction for extracting from the remote sensing image sample database is received, judging whether sample data meeting a first preset condition exists in the sample database according to the extraction requirement;
(3) Under the condition that sample data meeting a first preset condition does not exist in the remote sensing image sample library, receiving a label of a first object in the remote sensing image data, and extracting a second object in the remote sensing image data by using an iterative training and preset method according to the label so as to update the remote sensing image sample library;
(4) Under the condition that sample data meeting a first preset condition exists in the remote sensing image sample library, extracting the sample data in the remote sensing image sample library, performing model training, remote sensing image data prediction and optimization processing, extracting an increment sample according to a second preset rule, and updating the increment sample into the sample library.
In an optional manner, the establishing the remote sensing image sample library according to the first preset rule includes:
(S1) saving the remote sensing image sample library as a spatial database;
(S2) designing a data table structure in the spatial database as: image ID, image source, image geographic location, image band number, image resolution, imaging time, sample type, etc.;
the first preset rule is to store sample data according to a preset framing format or a first preset size data block according to requirements, the sample data stored according to standard framing is produced according to deep learning application requirements, and the first preset size data block stores the sample data for training by directly using the sample data.
In an optional manner, the determining, according to the extraction requirement, whether the sample data satisfying the first preset condition exists in the sample library includes:
judging whether sample data meeting the training requirement of the deep learning model exist in a sample library according to the sample types; the first preset condition is whether the quantity and the quality of sample data in the sample library meet the training requirement of the deep learning model.
In an optional manner, the receiving receives a label of a first object in the remote sensing image data, and according to the label, extracts a second object in the remote sensing image data by using an iterative training and preset method; comprising the following steps:
(T1) acquiring preprocessed remote sensing image data;
(T2) labeling a target object in target remote sensing image data in the remote sensing image data to obtain sample data;
(T3) performing model training according to the sample data to obtain a primary prediction model;
(T4) performing prediction and optimization processing on the remote sensing image data except the target remote sensing image data one by one, and training a primary prediction model by using the prediction sample data meeting the second preset condition and the corresponding remote sensing image data;
(T5) continuously repeating the step T4 until a third preset condition is met, so as to obtain an optimized prediction model;
(T6) predicting the remote sensing image by using the trained optimized prediction model to obtain sample marking data;
the first object is sample data obtained through labeling, and the second object is unlabeled sample data except the first object.
In one alternative, the preprocessing includes at least one of the following: orthographic correction, image fusion, bit depth conversion, image color homogenization and image mosaic.
In an alternative manner, the predicting and optimizing process trains the primary prediction model by using the prediction labeling data and the corresponding remote sensing image data meeting the second preset condition, including:
(M1) constructing a type rule set of the prediction sample data according to attribute information of the prediction sample data, wherein the attribute information comprises at least one of the following information: remote sensing indexes, sample shapes, structures, textures and sample space topological relations of the remote sensing image data corresponding to the predicted sample data;
(M2) calculating a confidence of the predicted sample data from the rule set;
(M3) screening the predicted sample data based on the confidence.
In an alternative manner, the method constructs a type rule set of the prediction sample data according to attribute information of the prediction sample data, and the method further includes: receiving first processing of the outline of the target object under the condition that the target object is a first type object, and deleting discrete points and holes in the predicted sample data; or, in the case that the target object is an object of the second class, receiving a second process on the outline of the target object.
In one alternative, the first process includes: profile erosion expansion operation and profile smoothing operation; the second process includes: profile erosion expansion operations and profile optimization operations.
In one alternative, the contour erosion expansion operation includes: performing convolution processing on the predicted sample data by using operators with characteristic sizes to finish the operation of the predicted sample; the contour smoothing operation includes: and disconnecting the connection of which the connection in the outline corresponding to the predicted sample is smaller than or equal to a first preset threshold value, and deleting the protruding part of which the protruding degree in the outline corresponding to the predicted sample is smaller than a second preset threshold value.
In an alternative, the method further comprises: and when the land use current situation vector diagram exists in the geographic area, adding the land use current situation vector diagram and corresponding remote sensing image data into a sample library.
In an alternative manner, the land use current situation vector diagram is a land use vector diagram within a preset time period, and the preset time period is a time period that is located before the time of acquiring the remote sensing image data and is within a preset interval from the time of acquiring the remote sensing image.
In an alternative manner, the extracting the sample data in the sample library includes:
when the sample data is first-class labeling sample data, cutting the remote sensing image data corresponding to the labeling sample data according to the geographic space position of the remote sensing image data corresponding to the labeling sample data, and generating remote sensing image blocks with the same size; the first type of marked sample data is marked sample data, wherein the distribution of target objects corresponding to the sample data in the remote sensing image data is smaller than or equal to a third preset threshold value; or alternatively, the process may be performed,
cutting remote sensing image data corresponding to the first labeling sample data according to the sequence from left to right and from top to bottom under the condition that the sample data is the second labeling sample data, and generating remote sensing image blocks with the same size; the second type of marked sample data is marked sample data, wherein the distribution of target objects corresponding to the marked sample data in the remote sensing image data is larger than or equal to a fourth preset threshold value.
In an optional manner, the extracting sample data in the remote sensing image sample library, performing model training, remote sensing image data prediction and optimization, extracting an incremental sample according to a second preset rule, and updating the incremental sample to the sample library, includes:
(N1) performing model training according to the sample data to obtain a prediction model;
(N2) predicting and optimizing the unlabeled remote sensing image by using the prediction model to obtain prediction sample data meeting a fourth preset condition;
and (N3) updating the sample data with the precision meeting the preset precision in the predicted sample data into a sample library.
The method for establishing and updating the remote sensing image sample library solves the problems of low labeling efficiency, inaccurate labeling result and unsatisfactory effect of manually labeling remote sensing image data in the related technology; the accuracy and the labeling efficiency of remote sensing image data labeling are improved, and the sample manufacturing and management efficiency is improved.
The foregoing description is only an overview of the present application, and is intended to be implemented in accordance with the teachings of the present application in order that the same may be more clearly understood and to make the same and other objects, features and advantages of the present application more readily apparent.
Drawings
Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the application. Also, like reference numerals are used to designate like parts throughout the figures.
FIG. 1 is a flowchart of a method for establishing and updating a remote sensing image sample library according to an embodiment of the present application;
FIG. 2 is a flowchart illustrating a method for creating and updating a remote sensing image sample library according to another embodiment of the present application;
fig. 3 is a flowchart of a method for creating and updating a remote sensing image sample library according to another embodiment of the present application.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present application more apparent, exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
In the description of embodiments of the present application, the terms "first," "second," and the like are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include at least one such feature. In the description of the present application, the meaning of "plurality" means at least two, for example, two, three, etc., unless specifically defined otherwise.
Fig. 1 is a flowchart of an implementation of a method for establishing and updating a remote sensing image sample library according to an embodiment of the present application.
Referring to fig. 1, the method for establishing and updating a remote sensing image sample library provided by the embodiment of the application can be specifically applied to intelligent terminals such as desktop computers, notebook computers, mainframe computers, tablet computers and the like; the method comprises the following steps:
step 101, when an operation instruction for establishing a remote sensing image sample library is received, establishing the remote sensing image sample library according to a first preset rule.
Specifically, in the embodiment of the application, a remote sensing image sample library is stored as a spatial database; the data table structure in the spatial database is designed as follows: image ID, image source, image geographic location, image band number, image resolution, imaging time, sample type, etc.; the first preset rule refers to that sample data is stored according to a preset framing format or a preset size data block according to requirements, sample data with any size and format can be produced according to deep learning application requirements according to standard framing stored sample data, the sample data can be directly used for training by storing the sample data with the specific size data block, and the sample data does not need to be cut.
Step 102, when an instruction is received from a remote sensing image sample database, judging whether sample data meeting a first preset condition exists in the sample database according to the extraction requirement.
Specifically, in the embodiment of the application, whether sample data meeting the training requirement of the deep learning model exists in a sample library is judged according to the sample types; the first preset condition is whether the quantity and the quality of sample data in the sample library meet the training requirement of the deep learning model.
And 103, under the condition that sample data meeting the first preset condition does not exist in the remote sensing image sample library, receiving the label of the first object in the remote sensing image data, and extracting a second object in the remote sensing image data by using an iterative training and preset method according to the label so as to update the remote sensing image sample library.
Specifically, in the embodiment of the application, the remote sensing image data after pretreatment is obtained. In the embodiment of the application, a manual annotation of a target object in target remote sensing image data in first remote sensing image data is received, and in some specific examples, the manual annotation refers to that a spatial position of a ground object of interest is sketched from the first remote sensing image data through professional annotation software (such as ARCGIS, MAPGIS and the like) and is stored in a vector markup file; specifically, in the embodiment of the present application, the target object in the remote sensing image data may be a specific feature in the remote sensing image data; for example, in some possible application scenarios, the target object may be an artificial feature such as a greenhouse, a windmill, an oil tank pipe, a golf course, or a natural feature such as a mountain, a wetland, or a lake in the first remote sensing image data. Specifically, in the embodiment of the present application, the target remote sensing image data in the remote sensing image data may be a small portion of remote sensing image data in the remote sensing image data.
Specifically, in the embodiment of the application, new sample data is extracted by using an iterative training and preset method, and model training is performed according to the sample data to obtain a primary prediction model; performing prediction optimization processing on the remote sensing image data except the target remote sensing image data one by one, and training a primary prediction model by using the prediction sample data meeting the second preset condition and the corresponding remote sensing image data; continuously repeating the steps until a third preset condition is met, and obtaining an optimized prediction model; and predicting the remote sensing image by using the trained optimized prediction model to obtain sample marking data.
Specifically, in the embodiment of the application, model training is performed according to sample data, and different models can be selected for training according to different ground object categories; for example, in some specific examples, a deep supervision significant (Deeply Supervised Salient, DSS) model or RCF model may be selected for extraction of samples of roads, water surfaces, buildings, etc., and in other possible examples, a Yolo model may be used for windmills, oil pipes, etc.
Specifically, in the embodiment of the application, the labeling data and the corresponding remote sensing image data can be stored in a remote sensing image sample library. Specifically, in the embodiment of the present application, the labeling sample data is vector labeling sample data stored in a vector format.
Step 104, under the condition that sample data meeting the first preset condition exists in the remote sensing image sample library, extracting the sample data in the remote sensing image sample library, performing model training, remote sensing image data prediction and optimization processing, extracting an incremental sample according to the second preset rule, and updating the incremental sample into the sample library.
Specifically, in the embodiment of the application, the remote sensing image data after pretreatment is obtained. Specifically, in the embodiment of the application, under the condition that the sample data is first-class labeling sample data, remote sensing image data corresponding to the labeling sample data are cut according to the geographic space position of the remote sensing image data corresponding to the labeling sample data, and remote sensing image blocks with the same size are generated; the class labeling sample data are labeling sample data, wherein the distribution of target objects corresponding to the sample data in the remote sensing image data is smaller than or equal to a third preset threshold value; or, under the condition that the sample data is the second type of marked sample data, cutting the remote sensing image data corresponding to the marked sample data according to the sequence from left to right and from top to bottom to generate remote sensing image blocks with the same size; the second type of marked sample data is marked sample data, wherein the distribution of target objects corresponding to the marked sample data in the remote sensing image data is larger than or equal to a fourth preset threshold value.
In some specific examples, the first type of labeling sample data may be sparse sample data, and in particular, the sparse sample data is sample data with sparse distribution in the remote sensing image data, for example, land features with less distribution, such as windmills, oil pipes, golf courses, and the like.
In some specific examples, the second type of labeling sample data may be non-sparse sample data, and in particular, the non-sparse sample data is sample data distributed densely in the remote sensing image data, such as ground features of roads, buildings, farmlands, and the like.
Specifically, in the embodiment of the present application, machine learning model training, remote sensing image data prediction and optimization processing are performed, and sample data is updated to a sample library, including: model training is carried out according to the sample data to obtain a prediction model; performing prediction optimization processing on unlabeled remote sensing images by using the prediction model to obtain prediction sample data meeting a fourth preset condition; and updating the increment sample data conforming to the second preset rule into a sample library.
Specifically, in the embodiment of the application, the labeling data and the corresponding remote sensing image data which accord with the precision can be updated to the remote sensing image sample database.
The remote sensing image sample library is established and used for storing sample data; judging whether sample data in a sample library meets preset conditions according to the model training requirement, if not, obtaining sample data by marking a small amount of target objects in the obtained remote sensing image data, extracting new sample data by using an iterative training and preset method, and updating the sample library; and if the preset condition is met, extracting sample data in the sample library, performing machine learning model training, remote sensing image data prediction and optimization processing, and updating the sample data into the sample library. The problems of low labeling efficiency, inaccurate labeling result and unsatisfactory effect of manually labeling remote sensing image data in the related technology are solved; the accuracy and the labeling efficiency of remote sensing image data labeling are improved, the efficiency of establishing a sample library is improved, and the labor cost is saved; the management of the sample database is facilitated, and the efficiency of extracting the marked sample vector data and the remote sensing image data from the sample database is improved
Fig. 2 is a flowchart of a method for creating and updating a remote sensing image sample library according to another embodiment of the present application.
Based on the foregoing embodiments, referring to fig. 2, a method for establishing and updating a remote sensing image sample library according to another embodiment of the present application includes the following steps:
in step 201, there are sample data in the remote sensing image sample library that do not satisfy the first preset condition.
Specifically, in the embodiment of the present application, the sample data in the sample library does not meet the requirement of model training, which may be that the sample data in the sample library does not exist or that the number and quality of the sample data do not meet the requirement.
Step 202, a small amount of labeling of a target object in the remote sensing image data is obtained, and sample data is obtained.
Specifically, in the embodiment of the present application, the pretreatment includes at least one of the following treatments: orthographic correction, image fusion, bit depth conversion, image color homogenization and image mosaic.
Specifically, in the embodiment of the application, a target object is taken as a greenhouse as a specific example for explanation, a small amount of marked greenhouse sample data is used for training a deep learning network, and the accuracy is ensured when the samples are marked. And cutting the remote sensing image data and the greenhouse vector marking data into 800-800 training data, marking the corresponding position of the greenhouse area in the training label as 255, and marking the value of the surrounding area as 0. The training data is amplified by rotation, translation, mirroring and other methods, and the sample data is increased.
And 203, performing model training according to the sample data to obtain a primary prediction model.
Specifically, in the embodiment of the present application, a primary prediction model of the greenhouse is obtained according to the vector marker data file of the small number of greenhouses marked in step 202 and the corresponding remote sensing image data training model. Specifically, in the embodiment of the instinct application, a DSS deep learning network is used for extracting a remote sensing image greenhouse, and the DSS model is improved on a VGGNet model, so that several short connections from deeper side output to shallower side output are added. In this way, higher level functionality may help reduce side output to better locate significant areas, while lower level functionality may help enrich higher level side output and finer detail. And inputting the initial greenhouse sample data after manual labeling into a DSS model for greenhouse feature extraction training, and learning and training to obtain a primary prediction model of the greenhouse.
And 204, performing predictive optimization processing on the remote sensing image data except the target remote sensing image data one by one, and training a primary predictive model by using the predictive sample data meeting the second preset condition and the corresponding remote sensing image data.
Specifically, in the embodiment of the present application, the currently processed remote sensing image data may be any remote sensing image data other than the target remote sensing image data; after the current processed remote sensing image data is predicted, judging whether the marking data for predicting and marking the current processed remote sensing image data through the primary prediction model meets a first preset condition or not so as to meet the predicted marking data of the first preset condition and the corresponding remote sensing image data to train the primary prediction model. Specifically, in the embodiment of the present application, the second preset condition may be a condition for screening the prediction annotation data, for example, in some specific examples, the prediction annotation result of the primary prediction model on the remote sensing image data may not be accurate, the prediction annotation data with inaccurate prediction needs to be removed, and the prediction annotation data with relatively accurate prediction is used as input data of model training to continuously train the prediction model, so as to improve the accuracy of the prediction model on the remote sensing image data processing. Specifically, in the embodiment of the present application, the second preset condition may be set according to the actual application requirement of the user, and in the embodiment of the present application, the second preset condition is not specifically limited.
Specifically, in the embodiment of the present application, under the condition that the predicted sample data meets the second preset condition, training the primary prediction model with the predicted sample data includes:
(1) Constructing a type rule set of the predicted sample data according to attribute information of the predicted sample data, wherein the attribute information comprises at least one of the following information: predicting remote sensing indexes, sample shapes, structures, textures and sample space topological relations of remote sensing image data corresponding to the sample data;
(2) Calculating confidence coefficient of the predicted sample data according to the rule set;
(3) Confidence screening predicts sample data.
In some alternatives, before constructing the type rule set of the prediction sample data according to the attribute information of the prediction sample data, the method further includes:
under the condition that the target object is a first type object, receiving first processing on the outline of the target object, and deleting discrete points and holes in remote sensing image data corresponding to the predicted sample data; alternatively, in the case where the target object is an object of the second class, a second process of the contour of the target object is received.
In some alternatives, the first process comprises: profile erosion expansion operation and profile smoothing operation; the second process includes: profile erosion expansion operation and profile optimization operation;
wherein the contour etch expansion operation comprises: and (5) convolving the predicted sample data by using operators with characteristic sizes to finish the operation of the predicted sample. Convolution processing predicts remote sensing image data corresponding to the sample data, and obtains local maximum and minimum values of the remote sensing image data; according to the local maximum value and the local minimum value, processing a prediction sample by open operation;
the contour smoothing operation includes: and disconnecting the connection of the outline corresponding to the predicted sample remote sensing image data with the connection less than or equal to a first preset threshold value, and deleting the protruding part with the protruding degree less than a second preset threshold value in the outline corresponding to the predicted sample remote sensing image data.
The contour optimization operation includes: and carrying out specific optimization operation on the target boundary according to the outline boundary characteristics of the second class of objects.
Specifically, in the embodiment of the present application, the first preset threshold and the second preset threshold may be set according to the actual processing requirement of the user on the remote sensing image data, and in the embodiment of the present application, the first preset threshold and the second preset threshold are not specifically limited.
Specifically, in the embodiment of the application, a trained primary prediction model is used for predicting other untagged remote sensing image areas to obtain a greenhouse prediction result. In some specific examples, the unlabeled image is cut into 800×800 predicted data, in order to ensure the boundary accuracy of the predicted result, a boundary of 50 pixels is overlapped between two adjacent pieces of data, all 800×800 remote sensing image data are predicted to obtain predicted result data, then all predicted result data are spliced into an image with the size before cutting, if the precision of the predicted result is higher, the predicted result is directly input into the DSS model for continuous training, and if the precision is lower, sample screening processing is needed.
In some specific examples, sample screening work is performed on the predicted result data of the greenhouse, including error extraction sample rejection and sample boundary optimization processing. Specifically, in the embodiment of the application, the extracted greenhouse results are subjected to error elimination by using the sample shape factors, the greenhouse shapes are more regular and are mostly rectangular, a rule set is constructed by utilizing the length-width ratio of the rectangle, and the extraction results which do not meet the conditions are eliminated. And then, sample boundary optimization is carried out by utilizing rules, and the sample boundary optimization can be carried out due to the fact that the greenhouse is regular in boundary comparison. Firstly, eliminating sample noise by using an expansion corrosion algorithm, wherein corrosion is carried out before expansion, so that an object area which does not contain structural elements can be deleted, the outline of the object is smoothed, narrow connection is disconnected, and a tiny protruding part is removed; then the extraction result is processed by utilizing the principle of regular boundary comparison of the greenhouse, firstly, the large obtuse angle and sharp angle in the interior angle of the polygon of the greenhouse are removed, the rectangular shape of the polygon is kept, then the edges which are possibly perpendicular or parallel to the polygon in the polygon are adjusted, finally, the positions of the corner points of the polygon are adjusted by extracting the corner points in the grid chart of the greenhouse result, and the positions of the boundary of the polygon are adjusted by linearly fitting the boundary of the polygon, so that the boundary of the polygon is more in line with the actual situation; if the precision of the screened greenhouse samples is not high, the precision of the samples can be improved through manual intervention treatment.
In some specific examples, the screened greenhouse sample prediction result data is input as a sample into the DSS model for continued training of the model. Specifically, firstly, sample preparation is carried out on the extracted greenhouse vector labeling data and the remote sensing image data corresponding to the extracted greenhouse vector labeling data to obtain sample data with the size of 800 x 800; and (3) keeping consistent with a sample format in initial model training, inputting greenhouse vector marking data with the size of 800 x 800 and corresponding remote sensing image data into a DSS network for training, predicting image data of other unlabeled areas by using a trained DSS model, if a prediction result is good, continuing training and learning, and if the prediction result is not ideal, using boundary optimization and manual intervention.
Step 205, repeating step 204 until a third preset condition is satisfied, thereby obtaining an optimized prediction model.
Specifically, in some specific examples, the third preset condition may be a number of loops of the predictive optimization loop process; in other specific examples, the third preset condition may also be an accuracy rate of the primary prediction model to predict the first remote sensing image data. In the embodiment of the present application, the specific form of the third preset condition is not limited either.
Specifically, in some specific examples, step 205 is repeated until the greenhouse accuracy predicted by the greenhouse prediction model meets the requirement, so as to obtain an optimized greenhouse prediction model.
And 206, predicting the remote sensing image by using the trained optimized prediction model to obtain sample marking data.
Specifically, in the embodiment of the application, the remote sensing image data marked by the optimized prediction model can be target remote sensing image data, and the target remote sensing image data is marked once by optimizing the preset model, so that the accuracy of marking the target object in the target remote sensing image data is improved; in some possible examples, the remote sensing image data marked by the optimized prediction model can also be other remote sensing image data except the target remote sensing image data, and the remote sensing image data is marked by the optimized prediction model, so that the marking efficiency and the marking accuracy of the remote sensing image data are improved.
In step 207, the sample data is updated into the sample library.
In the embodiment of the application, a primary prediction model is trained by marking a small amount of remote sensing image data; and predicting the remote sensing image data by using the primary prediction model, repeatedly training the primary prediction model according to the prediction result meeting the condition to obtain an optimized training model, marking the remote sensing image data by using the optimized training model, and updating the marking data and the corresponding remote sensing image data into a database. The problems of low labeling efficiency, inaccurate labeling result and unsatisfactory effect of manually labeling remote sensing image data in the related technology are solved; the accuracy and the labeling efficiency of remote sensing image data labeling are improved, the efficiency of establishing a sample library is improved, and the labor cost is saved.
Fig. 3 is a flowchart of a method for creating and updating a remote sensing image sample library according to another embodiment of the present application.
Based on the foregoing embodiments, referring to fig. 3, a method for establishing and updating a remote sensing image sample library according to another embodiment of the present application includes the following steps:
in step 301, sample data satisfying a first preset condition exists in a remote sensing image sample library.
Specifically, in the embodiment of the application, sample data in a sample library accords with the requirement of model training, namely, the quantity and the quality of the sample data accord with the requirement of model training.
Step 302, sample data in a sample library is extracted, and sample data is produced.
Specifically, in the embodiment of the present application, a target object is taken as a greenhouse as a specific example for explanation, in application of a greenhouse sample library, distribution of the greenhouse is relatively concentrated, as a sparse sample, labeled sample vector data corresponding to the greenhouse is extracted from the sample library, then image cutting is performed according to an area where the greenhouse labeled vector data is located, an image block with a size of 800 x 800 is generated, a relative 800 x 800 size mark file is generated, a greenhouse area value in a sample tag is 255, and values of other areas are 0.
And step 303, performing model training according to the sample data to obtain a prediction model.
Specifically, in the embodiment of the application, a prediction model of the greenhouse is obtained according to the vector marking data file of the greenhouse extracted from the sample library and the corresponding remote sensing image data training model.
And step 304, predicting and optimizing the remote sensing image data to obtain sample marking data.
Specifically, in the embodiment of the application, a greenhouse prediction model is used for predicting and optimizing unlabeled remote sensing images to obtain predicted sample data meeting a fourth preset condition; specifically, in the embodiment of the present application, the fourth preset condition is the same as the second preset condition in embodiment 2.
And step 305, extracting an incremental sample according to a second preset rule, and updating the incremental sample into the sample library.
Specifically, in the embodiment of the application, sample screening is performed on the predicted greenhouse result, and if the predicted result is good, the greenhouse extraction result can be directly updated into the database as a sample. Specifically, in some specific examples, the second preset rule means that the accuracy of the extracted sample meets the application requirement.
In the embodiment of the application, a prediction model is trained by extracting sample data in a sample library; and marking the remote sensing image data by using the prediction model, and updating the marked data and the corresponding remote sensing image data into a database. The problems of low labeling efficiency, inaccurate labeling result and unsatisfactory effect of manually labeling remote sensing image data in the related technology are solved; the accuracy and the labeling efficiency of remote sensing image data labeling are improved, the efficiency of establishing a sample library is improved, and the labor cost is saved.
Claims (11)
1. The method for establishing and updating the remote sensing image sample library is characterized by comprising the following steps of:
(1) When an operation instruction for establishing a remote sensing image sample library is received, establishing the remote sensing image sample library according to a first preset rule;
the first preset rule is to store sample data according to a preset framing format or a first preset size data block according to requirements, the sample data stored according to standard framing is sample data produced according to deep learning application requirements, and the first preset size data block stores the sample data for training by directly using the sample data;
(2) When an instruction for extracting from the remote sensing image sample database is received, judging whether sample data meeting a first preset condition exists in the sample database according to the extraction requirement;
judging whether sample data meeting a first preset condition exists in a sample library according to the extraction requirement, wherein the method comprises the following steps: judging whether sample data meeting the training requirement of the deep learning model exist in a sample library according to the sample types; the first preset condition is whether the quantity and the quality of sample data in the sample library meet the training requirement of the deep learning model;
(3) Under the condition that sample data meeting a first preset condition does not exist in the remote sensing image sample library, receiving a label of a first object in the remote sensing image data, and extracting a second object in the remote sensing image data by using an iterative training and preset method according to the label so as to update the remote sensing image sample library;
the method comprises the steps of receiving a label of a first object in the remote sensing image data, and extracting a second object in the remote sensing image data by using an iterative training and preset method according to the label; comprising the following steps:
(T1) acquiring preprocessed remote sensing image data;
(T2) labeling a target object in target remote sensing image data in the remote sensing image data to obtain sample data;
(T3) performing model training according to the sample data to obtain a primary prediction model;
(T4) predicting and optimizing the remote sensing image data except the target remote sensing image data one by one to obtain an optimized prediction model; predicting the remote sensing image by using the trained optimized prediction model to obtain sample marking data; the first object is sample data obtained through labeling, and the second object is unlabeled sample data except the first object;
(4) Under the condition that sample data meeting a first preset condition exists in the remote sensing image sample library, extracting the sample data in the remote sensing image sample library, performing model training, remote sensing image data prediction and optimization processing, extracting an increment sample according to a second preset rule, and updating the increment sample into the sample library.
2. The method of claim 1, wherein the establishing the remote sensing image sample library according to the first preset rule comprises:
(S1) saving the remote sensing image sample library as a spatial database;
(S2) designing a data table structure in the spatial database to be at least: image ID, image source, image geographic location, image band number, image resolution, imaging time, sample type.
3. The method of claim 1, wherein the step of determining the position of the substrate comprises,
training a primary prediction model by using the prediction sample data meeting the second preset condition and the corresponding remote sensing image data;
and continuously repeating the step T4 until a third preset condition is met.
4. The method of claim 3, wherein the predicting and optimizing process trains the primary prediction model with the prediction annotation data and the corresponding remote sensing image data satisfying the second preset condition, comprising:
(M1) constructing a type rule set of the prediction sample data according to attribute information of the prediction sample data, wherein the attribute information comprises at least one of the following information: remote sensing indexes, sample shapes, structures, textures and sample space topological relations of the remote sensing image data corresponding to the predicted sample data;
(M2) calculating a confidence of the predicted sample data from the rule set;
(M3) screening the predicted sample data based on the confidence.
5. The method of claim 4, wherein the constructing the set of type rules for the predicted sample data based on the attribute information for the predicted sample data, the method further comprises: receiving first processing of the outline of the target object under the condition that the target object is a first type object, and deleting discrete points and holes in the predicted sample data; or, in the case that the target object is an object of the second class, receiving a second process on the outline of the target object.
6. The method of claim 5, wherein the first processing comprises: profile erosion expansion operation and profile smoothing operation; the second process includes: profile erosion expansion operations and profile optimization operations.
7. The method of claim 6, wherein the contour etch expansion operation comprises: performing convolution processing on the predicted sample data by using operators with characteristic sizes to finish the operation of the predicted sample; the contour smoothing operation includes: and disconnecting the connection of which the connection in the outline corresponding to the predicted sample is smaller than or equal to a first preset threshold value, and deleting the protruding part of which the protruding degree in the outline corresponding to the predicted sample is smaller than a second preset threshold value.
8. The method according to claim 1, wherein the method further comprises: and when the land use current situation vector diagram exists in the geographic area, adding the land use current situation vector diagram and corresponding remote sensing image data into a sample library.
9. The method of claim 8, wherein the land use present vector map is a land use vector map within a preset time period that is a time period that is located before a time at which the remote sensing image data is acquired and that is within a preset interval from a time at which the remote sensing image data is acquired.
10. The method of claim 1, wherein the extracting sample data in the sample library comprises:
when the sample data is first-class labeling sample data, cutting the remote sensing image data corresponding to the labeling sample data according to the geographic space position of the remote sensing image data corresponding to the labeling sample data, and generating remote sensing image blocks with the same size; the first type of marked sample data is marked sample data, wherein the distribution of target objects corresponding to the sample data in the remote sensing image data is smaller than or equal to a third preset threshold value; or alternatively, the process may be performed,
cutting remote sensing image data corresponding to the labeling sample data according to the sequence from left to right and from top to bottom under the condition that the sample data is the second type labeling sample data, and generating remote sensing image blocks with the same size; the second type of marked sample data is marked sample data, wherein the distribution of target objects corresponding to the marked sample data in the remote sensing image data is larger than or equal to a fourth preset threshold value.
11. The method according to claim 1, wherein the extracting sample data in the remote sensing image sample library, performing model training, remote sensing image data prediction and optimization, extracting incremental samples according to a second preset rule, and updating the incremental samples into the sample library comprises:
(N1) performing model training according to the sample data to obtain a prediction model;
(N2) predicting and optimizing the unlabeled remote sensing image by using the prediction model to obtain prediction sample data meeting a fourth preset condition;
and (N3) updating the sample data with the precision meeting the preset precision in the predicted sample data into a sample library.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911372783.5A CN111144487B (en) | 2019-12-27 | 2019-12-27 | Method for establishing and updating remote sensing image sample library |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911372783.5A CN111144487B (en) | 2019-12-27 | 2019-12-27 | Method for establishing and updating remote sensing image sample library |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111144487A CN111144487A (en) | 2020-05-12 |
| CN111144487B true CN111144487B (en) | 2023-09-26 |
Family
ID=70521152
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911372783.5A Active CN111144487B (en) | 2019-12-27 | 2019-12-27 | Method for establishing and updating remote sensing image sample library |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111144487B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112884791B (en) * | 2021-02-02 | 2021-11-26 | 重庆市地理信息和遥感应用中心 | Method for constructing large-scale remote sensing image semantic segmentation model training sample set |
| CN113223042B (en) * | 2021-05-19 | 2021-11-05 | 自然资源部国土卫星遥感应用中心 | Intelligent acquisition method and equipment for remote sensing image deep learning sample |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105279519A (en) * | 2015-09-24 | 2016-01-27 | 四川航天系统工程研究所 | Remote sensing image water body extraction method and system based on cooperative training semi-supervised learning |
| CN108764263A (en) * | 2018-02-12 | 2018-11-06 | 北京佳格天地科技有限公司 | The atural object annotation equipment and method of remote sensing image |
| CN108921025A (en) * | 2018-06-01 | 2018-11-30 | 苏州中科天启遥感科技有限公司 | A kind of object level classification samples automatic selecting method of collaborative variation detection |
| CN109670060A (en) * | 2018-12-10 | 2019-04-23 | 北京航天泰坦科技股份有限公司 | A kind of remote sensing image semi-automation mask method based on deep learning |
| CN110298348A (en) * | 2019-06-12 | 2019-10-01 | 苏州中科天启遥感科技有限公司 | Remote sensing image building sample areas extracting method and system, storage medium, equipment |
| CN110309809A (en) * | 2019-07-09 | 2019-10-08 | 广西壮族自治区基础地理信息中心 | High Resolution Remote Sensing Satellites image sugarcane extracting method based on deep neural network |
| CN110502654A (en) * | 2019-08-26 | 2019-11-26 | 长光卫星技术有限公司 | A kind of object library generation system suitable for multi-source heterogeneous remotely-sensed data |
-
2019
- 2019-12-27 CN CN201911372783.5A patent/CN111144487B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105279519A (en) * | 2015-09-24 | 2016-01-27 | 四川航天系统工程研究所 | Remote sensing image water body extraction method and system based on cooperative training semi-supervised learning |
| CN108764263A (en) * | 2018-02-12 | 2018-11-06 | 北京佳格天地科技有限公司 | The atural object annotation equipment and method of remote sensing image |
| CN108921025A (en) * | 2018-06-01 | 2018-11-30 | 苏州中科天启遥感科技有限公司 | A kind of object level classification samples automatic selecting method of collaborative variation detection |
| CN109670060A (en) * | 2018-12-10 | 2019-04-23 | 北京航天泰坦科技股份有限公司 | A kind of remote sensing image semi-automation mask method based on deep learning |
| CN110298348A (en) * | 2019-06-12 | 2019-10-01 | 苏州中科天启遥感科技有限公司 | Remote sensing image building sample areas extracting method and system, storage medium, equipment |
| CN110309809A (en) * | 2019-07-09 | 2019-10-08 | 广西壮族自治区基础地理信息中心 | High Resolution Remote Sensing Satellites image sugarcane extracting method based on deep neural network |
| CN110502654A (en) * | 2019-08-26 | 2019-11-26 | 长光卫星技术有限公司 | A kind of object library generation system suitable for multi-source heterogeneous remotely-sensed data |
Non-Patent Citations (1)
| Title |
|---|
| 邵振峰 ; 张源 ; 黄昕 ; 朱秀丽 ; 吴亮 ; 万波 ; .基于多源高分辨率遥感影像的2 m不透水面一张图提取.武汉大学学报(信息科学版).2018,(第12期),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111144487A (en) | 2020-05-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102289991B (en) | Visual-variable-based automatic classification and configuration method of map lettering | |
| CN102136218B (en) | Automatic collocation method for map lettering during map making through computer | |
| CN111144487B (en) | Method for establishing and updating remote sensing image sample library | |
| CN108734120A (en) | Mark method, apparatus, equipment and the computer readable storage medium of image | |
| CN109635748B (en) | Method for extracting road characteristics in high-resolution image | |
| CN109849576B (en) | Method for assisting drawing by referring to gray level diagram | |
| CN111311746B (en) | Intelligent three-dimensional geological modeling method based on drilling data | |
| CN113223042B (en) | Intelligent acquisition method and equipment for remote sensing image deep learning sample | |
| Guilbert et al. | Terrain generalisation | |
| CN110688756B (en) | Collection and drawing integrated completion drawing manufacturing method | |
| CN109961129A (en) | A kind of Ocean stationary targets search scheme generation method based on improvement population | |
| CN108765445B (en) | Lung trachea segmentation method and device | |
| CN102609721B (en) | Remote sensing image clustering method | |
| CN112990001A (en) | Remote sensing image monomer building rapid and accurate extraction method applied to digital collection | |
| CN105894553A (en) | Street spatial form layout method based on grid selection | |
| CN115983385B (en) | Space environment entity construction method and device | |
| CN109166183B (en) | Anatomical landmark point identification method and identification equipment | |
| CN111414907A (en) | Data set labeling method, data set labeling device and computer-readable storage medium | |
| CN115546116A (en) | Method and system for extracting and calculating spacing of discontinuous surface of fully-covered rock mass | |
| CN104732292A (en) | Automatic planning method for main velocity line based on profile data | |
| CN115600732A (en) | City planning method and system based on big data analysis | |
| CN111598036B (en) | Urban group geographic environment knowledge base construction method and system of distributed architecture | |
| CN114565730A (en) | Method and device for processing map vector sample and computer readable storage medium | |
| Chen et al. | A new DEM generalization method based on watershed and tree structure | |
| CN113887737A (en) | Sample set automatic generation method based on machine 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 |