CN113158807A - Model self-training and optimizing system for remote sensing image - Google Patents
Model self-training and optimizing system for remote sensing image Download PDFInfo
- Publication number
- CN113158807A CN113158807A CN202110311388.7A CN202110311388A CN113158807A CN 113158807 A CN113158807 A CN 113158807A CN 202110311388 A CN202110311388 A CN 202110311388A CN 113158807 A CN113158807 A CN 113158807A
- Authority
- CN
- China
- Prior art keywords
- model
- training
- land
- remote sensing
- 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.)
- Granted
Links
- 238000012549 training Methods 0.000 title claims abstract description 57
- 238000001514 detection method Methods 0.000 claims abstract description 31
- 238000005457 optimization Methods 0.000 claims abstract description 27
- 230000011218 segmentation Effects 0.000 claims abstract description 27
- 230000008859 change Effects 0.000 claims abstract description 23
- 238000005516 engineering process Methods 0.000 claims abstract description 5
- 238000000034 method Methods 0.000 claims description 38
- 238000002372 labelling Methods 0.000 claims description 19
- 230000008569 process Effects 0.000 claims description 18
- 230000000694 effects Effects 0.000 claims description 13
- 238000010276 construction Methods 0.000 claims description 10
- 230000002452 interceptive effect Effects 0.000 claims description 9
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 9
- 238000012360 testing method Methods 0.000 claims description 7
- 238000011156 evaluation Methods 0.000 claims description 6
- 238000002474 experimental method Methods 0.000 claims description 6
- 238000007726 management method Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 6
- 239000002689 soil Substances 0.000 claims description 5
- 230000000007 visual effect Effects 0.000 claims description 5
- 238000013523 data management Methods 0.000 claims description 3
- 230000007812 deficiency Effects 0.000 claims description 3
- 238000011161 development Methods 0.000 claims description 3
- 230000018109 developmental process Effects 0.000 claims description 3
- 238000005065 mining Methods 0.000 claims description 3
- 210000002569 neuron Anatomy 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 238000012163 sequencing technique Methods 0.000 claims description 3
- 238000003860 storage Methods 0.000 claims description 3
- 101000878595 Arabidopsis thaliana Squalene synthase 1 Proteins 0.000 claims description 2
- 210000000988 bone and bone Anatomy 0.000 claims description 2
- 238000004519 manufacturing process Methods 0.000 abstract description 3
- 238000013527 convolutional neural network Methods 0.000 abstract 1
- 238000011160 research Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000005520 cutting process Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/54—Interprogram communication
- G06F9/547—Remote procedure calls [RPC]; Web services
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a model self-training and optimizing system of remote sensing images, which comprises a model training system, a model prediction system and a model optimizing system, wherein the model training system comprises a training platform, a model prediction platform and a model optimization platform; the model training system is used for enabling a user to independently select a preset model, creating a model meeting the self scene requirement, carrying out model training after uploading and marking pictures, and carrying out online prediction service after the training is finished; the model prediction system is based on a trained model, a user calls a related API (application program interface) interface, and remote sensing images are uploaded to segment and detect the ground objects in the area; the model optimization system is used for putting the detected remote sensing image as a training sample into model training again. The invention provides a target detection technology based on a deep convolutional neural network, solves the problems that the efficiency of manual identification and statistics is low and the requirement of social production efficiency improvement cannot be met in the prior art, provides an efficient means for a front-line operator to perform land type segmentation and statistics and extract change detection patterns, and is more worthy of popularization and application.
Description
Technical Field
The invention relates to the field of remote sensing image application, in particular to a model self-training and optimizing system for remote sensing images.
Background
The remote sensing image is a film or a photo for recording the size of electromagnetic waves of various ground objects, and is mainly divided into an aerial photo and a satellite photo. The remote sensing technology has the following advantages: (1) the method has the advantages of wide detection range and quick data acquisition, and can observe a large area from the air in a short time and obtain valuable remote sensing data from the large area; (2) the system can dynamically reflect the change of ground objects, and can periodically and repeatedly observe the ground of the same area, which is beneficial for people to find and dynamically track the change of a plurality of objects on the earth and research the change rule of the nature through the acquired remote sensing data; (3) the acquired data is comprehensive, a plurality of natural and human phenomena on the earth are comprehensively displayed, the shapes and the distribution of various things on the earth are macroscopically reflected, the characteristics of geology, landform, soil, vegetation, hydrology, artificial structures and other ground features are truly displayed, and the relevance among the geographic things is comprehensively revealed. And these data have the same presence in time.
The remote sensing image is mainly used in the fields of military reconnaissance, national soil resource investigation, city marking, traffic navigation and the like, and research targets comprise natural resources, airplanes, ships, vehicles, roads, ports, various buildings and the like. The current remote sensing image research mainly has the following difficulties: (1) the remote sensing image has large visual range, so that the image resolution is high, and the detection of a small target of the remote sensing image is difficult under a large viewpoint angle for a small-sized target in the image; (2) most targets in common images are horizontal, and when a remote sensing image is shot, the rotation invariance of the targets is also an important problem; (3) the background of the remotely sensed image is rather complex. Currently, remote sensing image detection based on deep learning is dedicated to solving the problems, and the invention provides a system capable of quickly and accurately establishing a self-training model and optimizing the model based on remote sensing images.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: the method aims to solve the problems that the remote sensing image has large image resolution due to large visual distance, and the detection of a small target of the remote sensing image is difficult under a large viewpoint angle for a small-sized target in the image; most targets in common images are horizontal, and when a remote sensing image is shot, the rotation invariance of the targets is also an important problem; the background of the remote sensing image is quite complex, so that a model self-training and optimizing system of the remote sensing image is provided.
The invention solves the technical problems through the following technical scheme, and the method comprises a model training system, a model prediction system and a model optimization system;
the model training system is used for enabling a user to independently select a preset model, creating a model meeting the self scene requirement, carrying out model training after uploading and marking pictures, and carrying out online prediction service after the training is finished;
the model prediction system is based on a trained model, a user calls a related API (application program interface) interface, and remote sensing images are uploaded to segment and detect the ground objects in the area;
the model optimization system is used for putting the detected remote sensing image as a training sample into model training again, namely from the training to the prediction of the model to the optimization of the model to form a complete product closed loop;
the model training system comprises the following steps:
s1: data acquisition:
aiming at three application directions of the remote sensing image: the method comprises the steps of oblique frame detection, land class segmentation and change detection, different data sets are established, in order to guarantee the representativeness and diversity of sample data, preset remote sensing images with five resolutions are respectively selected, the remote sensing images with preset areas in preset provincial regions are respectively selected according to provincial regions, and the generalization capability of a model is improved;
s2: the specific process of data annotation is as follows:
SS 1: the data marking of the oblique frame detection model supports a rectangular frame marking mode, and a rectangular frame is drawn according to a frame mark wrapped outside a marked object;
the labeling of the object includes: the method comprises the following steps of (1) counting 14 common statistical land features in cultivated land, woodland, grassland, greenhouse, gymnasium, railway station, airport, general building, park, pond, reservoir, natural lake, golf course and playground; uniformly selecting each type of ground object according to the image resolution in equal proportion;
SS 2: the data marking of the land type segmentation model supports multiple marking modes of polygon, broken line, smearing and interactive segmentation, and the land and land types are divided into cultivated land, garden land, forest land, grassland, business land, industrial and mining storage land, residential land, public management and public service land, special land, transportation land, water area and water conservancy facility land and other land according to preset rules;
SS3, data labeling of the change detection model supports multiple labeling modes of polygon, broken line, smearing and interactive segmentation, the labeling objects are mutual changes among cultivated land, grassland, water body, road, building, forest land, bare soil, manual piling and digging land, construction site and structure land, the labeling is carried out according to the structure of the building, when the image building in the front and back two stages has structural change, the labeling is carried out, and the labeling is not carried out due to the image change generated by the image tense under the condition that the land is not changed;
s3: constructing a model: abstracting and solving a high-dimensionality problem based on marked image data, constructing an implicit relation between features and label, expressing effects through parameter adjustment and characteristic optimization, and generating final data through fitting of a target function;
the contents of the model prediction system are as follows:
the evaluation employed by the model prediction system includes: the accuracy, the precision, the F value, the AUC and the NDCG are used for adjusting the parameters of the model through the evaluation indexes, model prediction is carried out by using a test set, and ABtest is carried out by using a large number of unmarked remote sensing images to verify the model;
in the ABtest process, the new model is always subjected to effect test by giving smaller flow initially due to uncertainty, and more flow is distributed when the effect is better than that of a base group model;
the specific contents of the model optimization system are as follows:
the model optimization process machine carries out the process of iterative upgrade on the model and the data, and the image factors comprise: basic data, construction characteristics, algorithm selection, experiment strategies, sequencing results and front position display;
in the whole process of applying the algorithm model, an ABTest experiment is utilized, a large amount of remote sensing data segmentation and detection results are utilized, manual optimization is carried out, the results are added into a training set, and the results are obtained through gradual training iteration.
Preferably, the online system comprises a visual interactive interface running at a computer end, a security management engine, a data management engine, a model training engine and a model application engine running at a Linux server end, the established algorithm model is trained through a GPU server, and the bottom development technologies are PaddleClass, PaddleDetection and PaddleSeg.
Preferably, in the step S3, the algorithm used in the model construction includes a method of early termination based on the loss value, a data enhancement method of performing operations such as flipping, rotating, clipping, deforming, and scaling on the remote sensing image, and a Dropout method of randomly disabling some neurons at a certain probability to avoid overfitting of the model.
Preferably, the network structure of the general semantic segmentation task adopted by the geo-class segmentation model includes: the deep LabV3+, the PSPNet and the OCRNet all adopt a method considering context information, and select fused OCR (Back bone HRNet-W64) and deep LabV3+ (back-bone ResNet101) as a final ground class segmentation model according to the mIoU score.
Preferably, the change detection model adopts an FPNRes-Unet model, the model is suitable for processing multi-scale problems of objects with different sizes in remote sensing images or remote sensing images with different resolutions, shallow detail information and deep semantic information of the images are fused, the model is based on Unet, and in order to enable the sizes of input images and output images to be the same, boundary filling is used after each convolution;
in the coding path, a residual error structure of ResNet18 is used for replacing all convolutional layers of Unet to extract image features, a branch path is added in the process of sampling at each level of a decoding path, FPN is fused into a network backbone of a model, the deficiency that the Unet model only outputs an original resolution layer is optimized through predicting classification results with different scales in the step of sampling, multi-scale information is utilized in back propagation and weight updating, feature maps of each layer are used for independent prediction, and the features of each layer are fully utilized for building change detection.
Compared with the prior art, the invention has the following advantages: this remote sensing image's model self-training and optimizing system, the target detection technique based on degree of depth convolution neural network has solved the artifical discernment statistical inefficiency of prior art, can't satisfy the demand that social production efficiency promoted, cuts apart and makes statistics for a ray of operation personnel ground class, and change detection pattern spot draws and provides an efficient means, provides quick timely decision-making for administrative department and provides technical support, has realized three big application directions of remote sensing image: the full-automatic process from data uploading and labeling to model training, testing and publishing of the inclined frame detection, the land classification segmentation and the change detection is free from any programming capability of a user, and is more convenient and worth of popularization and use.
Drawings
FIG. 1 is a block diagram of the system of the present invention.
Detailed Description
The following examples are given for the detailed implementation and specific operation of the present invention, but the scope of the present invention is not limited to the following examples.
As shown in fig. 1, the present embodiment provides a technical solution: a model self-training and optimizing system for remote sensing images comprises three parts, namely model training, model prediction and model optimization. The characteristics are described as follows:
the model training system allows a user to independently select a model meeting the self scene requirements on the basis of a preset model, model training is carried out after pictures are uploaded and labeled, and the online of prediction service can be realized after the training is finished;
the model prediction system is based on a trained model, a user calls a related API (application program interface), and remote sensing images are uploaded to complete the segmentation and detection of the ground features in the area;
the model optimization system is applied to putting detected remote sensing images as training samples into model training again, so that the accuracy of the model is continuously improved, namely, the model training, the prediction and the model optimization are carried out, and a complete product closed loop is formed. In this way, the user can combine the actual conditions of production to continually optimize the model until the model can meet the actual predicted needs of the user.
The model training system comprises the following steps:
s1: data acquisition
Aiming at three application directions of the remote sensing image: the method comprises the steps of oblique frame detection, land type segmentation and change detection, different data sets are established, in order to guarantee representativeness and diversity of sample data, remote sensing images with the resolution of 0.3m, 0.5m, 0.8m, 1m and 2m are selected respectively, and remote sensing images with the resolution of 23 kilometres in square are selected respectively according to province areas, wherein the remote sensing images are 23 kilometres in total, and the remote sensing images are selected from fifteen province areas such as Beijing city, Heilongjiang province, Jilin province, inner Mongolia autonomous region, Hebei province, Henan province and Hubei province, so that the generalization capability of a model is improved.
S2: data annotation
A. And the data marking of the oblique frame detection model supports a rectangular frame marking mode, and the rectangular frame is drawn according to the frame mark outsourced to the marked object. Annotating the object includes: the farmland, the woodland, the grassland, the greenhouse, the gymnasium, the railway station, the airport, the general building, the park, the pond, the reservoir, the natural lake, the golf course and the playground total 14 common statistical land features. Each type of ground object is uniformly selected according to the image resolution ratio in equal proportion.
B. The data marking of the land type segmentation model supports multiple marking modes of polygon, broken line, smearing and interactive segmentation, and the land types are divided into 12 types in total according to third national state survey technical regulation, such as cultivated land, garden land, forest land, grassland, business land, industrial and mining storage land, residential land, public management and public service land, special land, transportation land, water area and water conservancy facility land and other lands.
C. The data labeling of the change detection model supports multiple labeling modes of polygon, broken line, smearing and interactive segmentation, the labeling objects are mutual changes among 10 types of land, grassland, water, roads, buildings, forest lands, bare soil, manual piling and digging lands, construction sites and structures, and the labeling types are labeled. And for buildings, when the front and back image buildings have structural changes, the images also need to be marked. The image change caused by the image tense is not marked under the condition that the land type is not changed.
S3: model construction
The method is used for abstracting and solving the problem of high dimensionality based on marked image data, and aims to construct an implicit relation between features and label, and through parameter adjustment, feature optimization is carried out to pursue better effect expression. The minimum error and the best effect are pursued by fitting of the objective function.
The contents of the model prediction system are as follows:
the model prediction system employs evaluation indicators: accuracy, precision, F-number, AUC, NDCG. Parameters of the model are adjusted through the evaluation indexes to achieve the optimal offline effect, and besides model prediction is conducted through the test set, ABtest is conducted through a large number of unmarked remote sensing images to verify the quality of the model. In the ABtest process, the new model always gives smaller flow for effect test initially due to uncertainty, and if the effect is better than that of the base group model, more flow can be distributed to pursue the optimization of the overall effect on the line.
The contents of the model optimization system are as follows:
the model optimization process is a process of iteratively upgrading the model and the data, and the image factors comprise: basic data, construction characteristics, algorithm selection, experiment strategies, sequencing results and front position display. In the whole process of applying the algorithm model, each response action influences the effect expression of the model, except the important consideration of the algorithm and the characteristics, the influence of other non-data algorithms is introduced, the ABtest experiment is fully utilized, the segmentation and detection results of a large amount of remote sensing data are utilized, manual optimization is carried out, the results are added into a training set, and the model calling-in rate is improved through gradual training iteration.
The self-training and optimizing system operates on the basis of an online system, the online system comprises a visual interactive interface operating at a computer end, a security management engine, a data management engine, a model training engine and a model application engine operating at a Linux server end, the established algorithm model is trained through a GPU server, and the bottom development technologies are PaddleClass, PaddleDetection and PaddleSeg.
In the step S3 model construction, a (1) early termination method based on a loss value is adopted; (2) a data enhancement method for performing operations such as turning, rotating, cutting, deforming and zooming on the remote sensing image; (3) the Dropout approach, which randomly lets a part of neurons not work, is combined with a certain probability to avoid model overfitting.
The terrain segmentation model adopts three network structures with better performance in a general semantic segmentation task: DeepLabV3+, PSPNet and OCRNet. All adopt a method of considering context information. And selecting a fused OCR (backbone HRNet-W64) and DeepLabV3+ (backbone-bone ResNet101) as a final ground class segmentation model according to the mIoU score.
The change detection model adopts an FPNRes-Unet model, is suitable for processing multi-scale problems of objects with different sizes in remote sensing images or remote sensing images with different resolutions, and fully fuses shallow detail information and deep semantic information of the images. The model is based on Unet and, in order to make the input and output images the same size, boundary padding is used after each convolution. In the encoding path, a residual error structure of ResNet18 is used for replacing all convolutional layers of Unet to extract image characteristics, a branch path is added in the process of sampling at each level of a decoding path, FPN is fused into a network backbone of the model, and the deficiency that the Unet model only outputs to an original resolution layer is optimized by predicting classification results with different scales in the step of sampling. Therefore, the multi-scale information can be utilized in back propagation and weight updating, the feature maps of all layers are used for independent prediction, and the features of all layers are fully utilized for detecting the change of the building.
Furthermore, the terms "first", "second" and "first" 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 defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality" means at least two, e.g., two, three, etc., unless specifically limited otherwise.
In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.
Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.
Claims (5)
1. A model self-training and optimizing system of remote sensing images is characterized by comprising a model training system, a model prediction system and a model optimizing system;
the model training system is used for enabling a user to independently select a preset model, creating a model meeting the self scene requirement, carrying out model training after uploading and marking pictures, and carrying out online prediction service after the training is finished;
the model prediction system is based on a trained model, a user calls a related API (application program interface) interface, and remote sensing images are uploaded to segment and detect the ground objects in the area;
the model optimization system is used for putting the detected remote sensing image as a training sample into model training again, namely from the training to the prediction of the model to the optimization of the model to form a complete product closed loop;
the model training system comprises the following steps:
s1: data acquisition:
aiming at three application directions of the remote sensing image: the method comprises the steps of oblique frame detection, land class segmentation and change detection, different data sets are established, in order to guarantee the representativeness and diversity of sample data, preset remote sensing images with five resolutions are respectively selected, the remote sensing images with preset areas in preset provincial regions are respectively selected according to provincial regions, and the generalization capability of a model is improved;
s2: the specific process of data annotation is as follows:
SS 1: the data marking of the oblique frame detection model supports a rectangular frame marking mode, and a rectangular frame is drawn according to a frame mark wrapped outside a marked object;
the labeling of the object includes: the method comprises the following steps of (1) counting 14 common statistical land features in cultivated land, woodland, grassland, greenhouse, gymnasium, railway station, airport, general building, park, pond, reservoir, natural lake, golf course and playground; uniformly selecting each type of ground object according to the image resolution in equal proportion;
SS 2: the data marking of the land type segmentation model supports multiple marking modes of polygon, broken line, smearing and interactive segmentation, and the land and land types are divided into cultivated land, garden land, forest land, grassland, business land, industrial and mining storage land, residential land, public management and public service land, special land, transportation land, water area and water conservancy facility land and other land according to preset rules;
SS3, data labeling of the change detection model supports multiple labeling modes of polygon, broken line, smearing and interactive segmentation, the labeling objects are mutual changes among cultivated land, grassland, water body, road, building, forest land, bare soil, manual piling and digging land, construction site and structure land, the labeling is carried out according to the structure of the building, when the image building in the front and back two stages has structural change, the labeling is carried out, and the labeling is not carried out due to the image change generated by the image tense under the condition that the land is not changed;
s3: constructing a model: abstracting and solving a high-dimensionality problem based on marked image data, constructing an implicit relation between features and label, expressing effects through parameter adjustment and characteristic optimization, and generating final data through fitting of a target function;
the contents of the model prediction system are as follows:
the evaluation employed by the model prediction system includes: the accuracy, the precision, the F value, the AUC and the NDCG are used for adjusting the parameters of the model through the evaluation indexes, model prediction is carried out by using a test set, and ABtest is carried out by using a large number of unmarked remote sensing images to verify the model;
in the ABtest process, the new model is always subjected to effect test by giving smaller flow initially due to uncertainty, and more flow is distributed when the effect is better than that of a base group model;
the specific contents of the model optimization system are as follows:
the model optimization process machine carries out the process of iterative upgrade on the model and the data, and the image factors comprise: basic data, construction characteristics, algorithm selection, experiment strategies, sequencing results and front position display;
in the whole process of applying the algorithm model, an ABTest experiment is utilized, a large amount of remote sensing data segmentation and detection results are utilized, manual optimization is carried out, the results are added into a training set, and the results are obtained through gradual training iteration.
2. The model self-training and optimization system for remote sensing images of claim 1, wherein: the self-training and optimizing system operates on the basis of an online system, the online system comprises a visual interactive interface operating at a computer end, a security management engine, a data management engine, a model training engine and a model application engine operating at a Linux server end, the established algorithm model is trained through a GPU server, and the bottom development technologies are PaddleClass, PaddleDetection and PaddleSeg.
3. The model self-training and optimization system for remote sensing images of claim 1, wherein: in the step S3, in the model construction, the algorithm includes a loss value-based early termination method, a data enhancement method for performing operations such as flipping, rotating, clipping, deforming, scaling, and the like on the remote sensing image, and a Dropout method for randomly disabling a part of neurons at a certain probability to avoid overfitting of the model.
4. The model self-training and optimization system for remote sensing images of claim 1, wherein: the network structure of the general semantic segmentation task adopted by the ground class segmentation model comprises the following steps: the deep LabV3+, the PSPNet and the OCRNet all adopt a method considering context information, and select fused OCR (Back bone HRNet-W64) and deep LabV3+ (back-bone ResNet101) as a final ground class segmentation model according to the mIoU score.
5. The model self-training and optimization system for remote sensing images of claim 1, wherein: the change detection model adopts an FPNRes-Unet model, is suitable for processing objects with different sizes in remote sensing images or the multi-scale problem of remote sensing images with different resolutions, fuses shallow detail information and deep semantic information of the images, is based on Unet, and uses boundary filling after each convolution in order to ensure that the sizes of input and output images are the same;
in the coding path, a residual error structure of ResNet18 is used for replacing all convolutional layers of Unet to extract image features, a branch path is added in the process of sampling at each level of a decoding path, FPN is fused into a network backbone of a model, the deficiency that the Unet model only outputs an original resolution layer is optimized through predicting classification results with different scales in the step of sampling, multi-scale information is utilized in back propagation and weight updating, feature maps of each layer are used for independent prediction, and the features of each layer are fully utilized for building change detection.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110311388.7A CN113158807B (en) | 2021-03-24 | 2021-03-24 | Model self-training and optimizing system for remote sensing image |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110311388.7A CN113158807B (en) | 2021-03-24 | 2021-03-24 | Model self-training and optimizing system for remote sensing image |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113158807A true CN113158807A (en) | 2021-07-23 |
CN113158807B CN113158807B (en) | 2024-02-09 |
Family
ID=76888249
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110311388.7A Active CN113158807B (en) | 2021-03-24 | 2021-03-24 | Model self-training and optimizing system for remote sensing image |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113158807B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113762222A (en) * | 2021-11-08 | 2021-12-07 | 阿里巴巴达摩院(杭州)科技有限公司 | Method and device for processing surface feature elements, storage medium and processor |
CN113807301A (en) * | 2021-09-26 | 2021-12-17 | 武汉汉达瑞科技有限公司 | Automatic extraction method and automatic extraction system for newly-added construction land |
CN113989660A (en) * | 2021-10-14 | 2022-01-28 | 浙江数维科技有限公司 | Method for detecting different time phase image changes |
CN117876840A (en) * | 2023-11-30 | 2024-04-12 | 中国科学院空天信息创新研究院 | Remote sensing basic model rapid training method and system based on template editing |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111161218A (en) * | 2019-12-10 | 2020-05-15 | 核工业北京地质研究院 | High-resolution remote sensing image change detection method based on twin convolutional neural network |
CN111582043A (en) * | 2020-04-15 | 2020-08-25 | 电子科技大学 | High-resolution remote sensing image ground object change detection method based on multitask learning |
WO2020244261A1 (en) * | 2019-06-05 | 2020-12-10 | 中国科学院长春光学精密机械与物理研究所 | Scene recognition system for high-resolution remote sensing image, and model generation method |
-
2021
- 2021-03-24 CN CN202110311388.7A patent/CN113158807B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020244261A1 (en) * | 2019-06-05 | 2020-12-10 | 中国科学院长春光学精密机械与物理研究所 | Scene recognition system for high-resolution remote sensing image, and model generation method |
CN111161218A (en) * | 2019-12-10 | 2020-05-15 | 核工业北京地质研究院 | High-resolution remote sensing image change detection method based on twin convolutional neural network |
CN111582043A (en) * | 2020-04-15 | 2020-08-25 | 电子科技大学 | High-resolution remote sensing image ground object change detection method based on multitask learning |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113807301A (en) * | 2021-09-26 | 2021-12-17 | 武汉汉达瑞科技有限公司 | Automatic extraction method and automatic extraction system for newly-added construction land |
CN113807301B (en) * | 2021-09-26 | 2024-06-07 | 武汉汉达瑞科技有限公司 | Automatic extraction method and automatic extraction system for newly-added construction land |
CN113989660A (en) * | 2021-10-14 | 2022-01-28 | 浙江数维科技有限公司 | Method for detecting different time phase image changes |
CN113762222A (en) * | 2021-11-08 | 2021-12-07 | 阿里巴巴达摩院(杭州)科技有限公司 | Method and device for processing surface feature elements, storage medium and processor |
CN117876840A (en) * | 2023-11-30 | 2024-04-12 | 中国科学院空天信息创新研究院 | Remote sensing basic model rapid training method and system based on template editing |
Also Published As
Publication number | Publication date |
---|---|
CN113158807B (en) | 2024-02-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113158807B (en) | Model self-training and optimizing system for remote sensing image | |
Li et al. | RSI-CB: A large scale remote sensing image classification benchmark via crowdsource data | |
Chen et al. | Global land cover mapping at 30 m resolution: A POK-based operational approach | |
Kalantar et al. | Drone-based land-cover mapping using a fuzzy unordered rule induction algorithm integrated into object-based image analysis | |
Klouček et al. | How does data accuracy influence the reliability of digital viewshed models? A case study with wind turbines | |
Jagannathan et al. | Deep learning for the prediction and classification of land use and land cover changes using deep convolutional neural network | |
CN102054166B (en) | A kind of scene recognition method for Outdoor Augmented Reality System newly | |
CN107918776A (en) | A kind of plan for land method, system and electronic equipment based on machine vision | |
CN113223042B (en) | Intelligent acquisition method and equipment for remote sensing image deep learning sample | |
Li et al. | Extracting check dam areas from high‐resolution imagery based on the integration of object‐based image analysis and deep learning | |
Liu et al. | Land use and land cover mapping in China using multimodal fine-grained dual network | |
Canedo et al. | Uncovering archaeological sites in airborne LiDAR data with data-centric artificial intelligence | |
Comert et al. | Rapid mapping of forested landslide from ultra-high resolution unmanned aerial vehicle data | |
Raju et al. | Object Recognition in Remote Sensing Images Based on Modified Backpropagation Neural Network. | |
CN107908858B (en) | Coastal wetland waterfowl movement path simulation method based on GIS and RS | |
CN106339423A (en) | Method and device for dynamically updating sugarcane planting information | |
Zhou | GeoAI-Enhanced Techniques to Support Geographical Knowledge Discovery from Big Geospatial Data | |
Zhao et al. | Remote sensing image surface feature classification based on VGG-UNet | |
Hussaini et al. | A review of water level fluctuation models and modelling initiatives | |
Abdallah et al. | Olive tree and shadow instance segmentation based on Detectron2 | |
Wang et al. | Interactive image segmentation based field boundary perception method and software for autonomous agricultural machinery path planning | |
CN116189000B (en) | Land parcel analysis method and apparatus | |
Quintus et al. | Evaluating Mask R‐CNN models to extract terracing across oceanic high islands: A case study from Sāmoa | |
Matkar et al. | Land use/land cover changes patterns using geospatial techniques-Satara district, Maharastra, India: a case study | |
Schrum | Computational Geometry Applied to Innovations in GIS |
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 |