CN109961105A - A kind of Classification of High Resolution Satellite Images method based on multitask deep learning - Google Patents
A kind of Classification of High Resolution Satellite Images method based on multitask deep learning Download PDFInfo
- Publication number
- CN109961105A CN109961105A CN201910274699.3A CN201910274699A CN109961105A CN 109961105 A CN109961105 A CN 109961105A CN 201910274699 A CN201910274699 A CN 201910274699A CN 109961105 A CN109961105 A CN 109961105A
- Authority
- CN
- China
- Prior art keywords
- network
- sampled
- image
- sample
- classification
- 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
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/20—Image enhancement or restoration by the use of local operators
- G06T5/30—Erosion or dilatation, e.g. thinning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/13—Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/176—Urban or other man-made structures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30181—Earth observation
- G06T2207/30184—Infrastructure
Abstract
The invention discloses a kind of Classification of High Resolution Satellite Images methods based on multitask deep learning, calculate NDVI and NDWI first with four wave band images, and carry out the sample production of high-resolution remote sensing image based on this;By the screening to sample, training data is divided into complex samples and simple sample;The deep learning network for constructing multitask simultaneously, is first trained the network of complex samples, then freezes the partial parameters, be trained to the network of simple sample, obtains the disaggregated model that building and road are reinforced;Then the classification that test data is carried out with the model, obtains classification results;Edge detection and conditional opening operation finally are carried out to classification results, remove hole.The present invention improves the accuracy of identification of building and road, shortens the training time, optimizes classification results, can be used for the application problems such as earth's surface information extraction and variation detection.
Description
Technical field
The invention belongs to remote sensing and photogrammetric technology field, are related to a kind of deep learning method of multitask, especially
It is related to a kind of terrain classification method that high-resolution urban remote sensing image is directed to based on multitask deep learning.
Background technique
The mechanized classification of high-resolution remote sensing image is the analysis of satellite image data and a critical issue in application.
With the high speed development of remote sensing technology, image spatial resolution has reached higher level, and the increasingly maturation of machine learning is also
Image automatic identification and classification bring new opportunity.Automation image classification facilitates efficiently using image data, greatly
Width improves the working efficiency of previous human interpretation, improves accuracy of identification, in necks such as building extraction, road extraction, variation detections
There is good application prospect in domain.
High-resolution urban remote sensing image is abundant comprising terrestrial object information, and atural object complexity is high, and main at present automatic point
Class method has markov random file, condition random field, SVM, decision tree etc., but since these methods depend on image feature
Selection, it is difficult to be extended to the image classification of large area.Deep learning network efficiently solves the problems, such as Feature Selection, passes through
The automatic study of depth network model, can obtain high-level semantic feature, obtain more accurate semantic segmentation result.
Semantic segmentation based on deep learning is to carry out Pixel-level to image to classify end to end, available image classification
It is relatively fine as a result, still due to the particularity of high-resolution remote sensing image, the image classification of Pixel-level is often obtained more
Broken classification results cannot meet the broken situation of the genuine property of atural object, especially building and road, so that network point
The result of class cannot be directly used to produce.It is limited by the image size of network query function, the terrestrial object information for including is also very limited, each shadow
The complexity of picture is different, is all deep learning network problems faced in the automatic classification application of remote sensing image.
Summary of the invention
The present invention mainly solves the prior art in the classification of high-resolution remote sensing image, building and classification of road knot
Fruit is more broken, and erroneous detection and all higher problem of omission factor: providing a kind of deep learning network of optimization, can be for difference
The complexity of scene carries out the adjustment of e-learning, and is optimized by marginal information to testing result, can be effectively improved
The precision of terrain classification, while improving network parameter training effectiveness.
The technical scheme adopted by the invention is that:
A kind of high-resolution remote sensing image building extracting method based on multitask deep learning, which is characterized in that packet
Include following steps:
One pretreatment
Step 1: selection remote sensing image region utilizes four wave bands (red, green, blue, near-infrared) high-resolution for making sample
Rate urban remote sensing image carries out wave band calculating, right by NDWI (normalization aqua index), NDVI (normalized differential vegetation index) first
Water body and vegetation carry out rough sort;
Step 2: continue to delineate the sample of building and road on the basis of step 1, carries out artificial sample optimization, thus
To the calibration sample class of building, road, water body, vegetation and other these fifth types;
Step 3: by DSM (Digital Surface Model, the numerical cutting tool) number in the remote sensing image region of selection
According to as a wave band, is merged with four wave band images in step 1, obtain five wave band image samples;
Step 4: by five wave band image samples of step 3 it is consistent with the calibration sample in step 2 be cut to be suitble to network meter
The size of calculation (in the present invention by taking 500*500 as an example);
Step 5: the image sample after cutting in step 4 is divided into simple scenario, complicated field by the calibration sample according to step 2
Two class sample data of scape (in each sample, in any road, vegetation or water body a kind of accounting be more than 90% be considered as it is simple
Otherwise scene sample data is considered as complex scene sample data);
Two building deep learning networks
Step 6: refer to 2 structure of attached drawing, construct the deep learning network of multitask, network be divided into two down-sampled parts,
One rises sampling section and a multitask network portion, and the down-sampled part includes " the first down-sampled part ", " the second drop
Sampling section " constitutes " network 2 " by the second down-sampled part, liter sampling section, is suitable for complex scene;It is down-sampled by first
Partially, it rises sampling section and constitutes " network 1 ", be suitable for simple scenario;The multitask network portion constitutes " network 3 ".
The specific implementation of step 6 is as follows:
Step 6.1: constructing " network 2 " of complex scene, the learning difficulty of complex scene is higher, using resnet50 convolution
Neural network structure is trained, using the down-sampled structure of five layers of network;It rises in sampling process, thinks in conjunction with the design of Unet
Road is finally reduced to the image format of 500*500*64 to each layer of progress deconvolution;
Step 6.2: constructing " network 1 " of simple scenario, the network parameter that simple scenario needs is less, down-sampled by one
Sampling section is risen with one to constitute;Down-sampled part uses the structure of five layer networks, every layer network include two 3*3 convolutional layers with
It is one relu layers, down-sampled by max_pooling layers of progress between every layer, it is down-sampled by four layers, by the shadow of 500*500*5
Then format as becoming 32*32*1024 carries out a liter sampling with two dimension deconvolution, the output result for rising sampling is 500*500*
64;
Step 6.3: all liters of sampled results of step 6.1 and step 6.2 being all input to " network 3 ", for carrying out three
The study respectively of kind task;" network 3 " is mainly using 3*3 convolutional layer and relu layers of composition, input data format 500*500*
64, main task study in, output data format 500*500*5 is divided into five kinds of classifications, as a classification results;Two
In secondary task study, output data format 500*500*2, two secondary tasks are two points to road and building respectively
Class, as other two classification results.
In above-mentioned depth network structure, the processing of simple sample data " network 1 " therethrough+" network 3 ", and complicated sample
Notebook data is handled by " network 2 "+" network 3 ", and " network 1 ", " network 2 ", which share, rises sampling section, arrow in network
Indicate the flow direction of sample data.
Step 6.4: three classification results in step 6.3 being subjected to the calculating of loss value, main task weight is set as 1, and two auxiliary
The weight of task is all set as 0.4, final loss=loss1*2+loss2*0.4+loss3*0.4, and carries out to loss value
The calculating of AdamOptimizer is trained network parameter by iteration.
Three training
Step 7: by complex scene image sample obtained in step 5 and its corresponding calibration sample the input depth
Learning network, the training to the wherein second down-sampled part, the part progress parameter for rising sampling section and multitask network;It completes
After training, the training of freezing step 7 obtains model parameter, prepares for step 8 training step;
Step 8: by simple scenario image sample obtained in step 5 and its corresponding calibration sample the input depth
Learning network carries out the training of the first down-sampled network parameter, by the ginseng for rising sampling and multitask trained in step 7
Number freezes, the model parameter for only the first down-sampled network portion for needing to learn.
Four classification and post-processing
Step 9: carrying out the calculating of scene information entropy and variance to test section image, be divided into complex scene and letter with threshold method
Single game scape, and carry out ground to the test zone high-resolution urban remote sensing image of different complexities with the model parameter that training obtains
Object classification;
Step 10: it is down-sampled to the high-resolution remote sensing image progress of test section, then using canny operator to down-sampled
As a result the extraction of marginal information is carried out;
Step 11: the marginal information obtained according to step 10 handles the broken figure spot of the classification results of step 9,
Obtain more complete terrain classification result.
The specific implementation of step 10 is as follows:
Step 10.1: first by the high resolution image of 0.1m it is down-sampled be 0.5m, then to the image after down-sampled into
Row GaussianBlur (Gaussian smoothing) inhibits false edge;
Step 10.2: carrying out the edge detection of canny operator, obtain edge detection results.
The specific implementation of step 11 is as follows:
Step 11.1: the classification results of step 9 being carried out with the expansion of having ready conditions of edge constraint, by the resulting edge of step 10
Testing result expands the classification results of step 9 as exposure mask, if there is edge in the region, no in expansion process
Carry out expansion calculating.
Step 11.2: classification results being carried out with the corrosion of having ready conditions of edge constraint, equally examines the resulting edge of step 10
Fruit come to an end as exposure mask, the expansion results that step 11.1 obtains are corroded, carries out corruption at pixel existing for edge for having
Erosion processing.
The deep learning network that the present invention designs, can be effectively improved the precision of terrain classification, while improve network parameter
Training effectiveness.
Detailed description of the invention
Fig. 1 is the overview flow chart of the embodiment of the present invention;
Fig. 2 is the multitask network design schematic diagram of the embodiment of the present invention;
Fig. 3 is complex scene and simple scenario schematic diagram;
Fig. 4 is the Canny operator edge detection figure of the embodiment of the present invention;
Fig. 5 is that the expansion of having ready conditions of the embodiment of the present invention corrodes schematic diagram with having ready conditions.
Specific embodiment
Understand for the ease of those of ordinary skill in the art and implement the present invention, with reference to the accompanying drawings and embodiments to this hair
It is bright to be described in further detail, it should be understood that implementation example described herein is merely to illustrate and explain the present invention, not
For limiting the present invention.
Referring to Fig.1, the present invention provides a kind of Classification of High Resolution Satellite Images method based on multitask deep learning, packet
Include following steps:
Step 1: carrying out wave band calculating using four wave bands (red, green, blue, near-infrared) high-resolution urban remote sensing image, lead to
It crosses NDWI, NDVI and rough sort is carried out to water body and vegetation,
Step 2: delineate the sample of building and road on the basis of step 1, carry out artificial sample optimization, obtain building,
Road, water body, vegetation and other these fifth types calibration sample class;
Step 3: by DSM (Digital Surface Model, the numerical cutting tool) data in the region as a wave
Section, merges with four wave band images in step 1, obtains five wave band image samples,
Step 4: the calibration sample in the five wave band image samples and step 2 of step 3 is all cut to unified be suitable for
The size of network query function (present invention is by taking the size of 500*500 as an example);
Step 5: the image of step 4 being divided into simple scenario, two class of complex scene, each sample according to calibration sample
In, a kind of accounting is more than 90% to be considered as simple scenario in any road, vegetation or water body, is otherwise considered as complex scene).Figure
3 be complex scene and simple scenario schematic diagram.
Step 6.1: see Fig. 2.The convolutional Neural " network 2 " of complex scene is constructed, the learning difficulty of complex scene is higher,
So being trained using resnet50 network structure, using the down-sampled structure of five layers of network;It rises in sampling process, in conjunction with
The mentality of designing of Unet is finally reduced to the image format of 500*500*64 to each layer of progress deconvolution;
Step 6.2: constructing the convolutional Neural " network 1 " of simple scenario, the network parameter that simple scenario needs is less, uses
The structure of five layer networks, every layer network includes two 3*3 convolutional layers and one relu layers, by max_pooling layers between every layer
Carry out it is down-sampled, it is down-sampled by four layers, the image of 500*500*5 is become to the format of 32*32*1024, then with two dimension
Deconvolution carries out a liter sampling, rises the output result of sampling for 500*500*64, only to the down-sampled parameter of network 1 in training process
It is trained, rises the training result that sampling section also uses step 6.1;
Step 6.3: all liters of sampled results of step 6.1 and step 6.2 being all input to " network 3 ", three kinds is carried out and appoints
The study respectively of business, network 3 is mainly with 3*3 convolutional layer and relu layers of composition, input data format 500*500*64, main task
In study, output data format 500*500*5 is divided into five kinds of classifications, as a classification results;In secondary task study,
Output data format is 500*500*2, and two secondary tasks are two classification to road and building respectively, as other two point
Class result.
Step 6.4: three classification results in step 6.3 being subjected to the calculating of loss value with sample respectively, main task weight is set
It is 1, the weight of two secondary tasks is all set as 0.4, final loss=loss1*2+loss2*0.4+loss3*0.4, and to loss
Value carries out the calculating of AdamOptimizer, is trained by iteration to network parameter.
Step 7: by the input of complex scene image sample obtained in the step 5 deep learning network, to wherein the
The part of two down-sampled parts, liter sampling section and multitask network carries out the training of parameter;After completing training, freezing step 7
Training obtains model parameter, prepares for step 8 training step;
Step 8: simple scenario image sample obtained in step 5 being inputted into deep learning network, it is down-sampled to carry out first
The training of network parameter is only involved in simple sample based on the liter sampling and multitask parameter freezed trained in step 7
This calculating, but without the adjustment of parameter in training, what this step needed to learn only has the first down-sampled network portion
Parameter;
Step 9: see Fig. 3.The calculating that scene information entropy and variance are carried out to test section image, is divided into complexity with threshold method
Scene and simple scenario, and with trained obtained model parameter to the test zone high-resolution urban remote sensing shadow of different complexities
As carrying out terrain classification.
Step 10.1: see Fig. 4.First by the high resolution image of 0.1m it is down-sampled be 0.5m, then to after down-sampled
Image carry out GaussianBlur (Gaussian smoothing), inhibit false edge;
Step 10.2: carrying out the edge detection of canny operator to the result of step 10.1, obtain edge detection results.Fig. 4
For Canny operator edge detection result.
Step 11: the marginal information obtained according to step 10 handles the broken figure spot of the classification results of step 9,
Obtain more complete terrain classification result.
As shown in figure 5, the specific implementation of step 11 is as follows:
Step 11.1: the classification results of step 9 being carried out with the expansion of having ready conditions of edge constraint, by the resulting edge of step 10
Testing result expands the classification results of step 9 as exposure mask, if there is edge in the region, no in expansion process
Carry out expansion calculating.
Step 11.2: classification results being carried out with the corrosion of having ready conditions of edge constraint, equally examines the resulting edge of step 10
Fruit come to an end as exposure mask, the expansion results that step 11.1 obtains are corroded, carries out corruption at pixel existing for edge for having
Erosion processing.
It should be understood that the part that this specification does not elaborate belongs to the prior art.
It should be understood that the above-mentioned description for preferred embodiment is more detailed, can not therefore be considered to this
The limitation of invention patent protection range, those skilled in the art under the inspiration of the present invention, are not departing from power of the present invention
Benefit requires to make replacement or deformation under protected ambit, fall within the scope of protection of the present invention, this hair
It is bright range is claimed to be determined by the appended claims.
Claims (3)
1. a kind of high-resolution remote sensing image building extracting method based on multitask deep learning, which is characterized in that including
Following steps:
One pretreatment
Step 1: selection remote sensing image region utilizes four wave bands (red, green, blue, near-infrared) high-resolution city for making sample
City's remote sensing image carries out wave band calculating, first by NDWI (normalization aqua index), NDVI (normalized differential vegetation index) to water body
Rough sort is carried out with vegetation;
Step 2: continuing to delineate the sample of building and road on the basis of step 1, artificial sample optimization is carried out, to be built
Build the calibration sample class of object, road, water body, vegetation and other these fifth types;
Step 3: DSM (Digital Surface Model, the numerical cutting tool) data in the remote sensing image region of selection are made
It for a wave band, is merged with four wave band images in step 1, obtains five wave band image samples;
Step 4: being cut to be suitble to network query function by five wave band image samples of step 3 are with the calibration sample in step 2 consistent
Size;
Step 5: the image sample after cutting in step 4 is divided into simple scenario, complex scene two by the calibration sample according to step 2
Class sample data;
Two building deep learning networks
Step 6: construct the deep learning network of multitask, network be divided into two down-sampled parts, one rise sampling section and one
A multitask network portion, the down-sampled part includes " the first down-sampled part ", " the second down-sampled part ", by the second drop
Sampling section rises sampling section composition " network 2 ", is suitable for complex scene;It is made of the first down-sampled part, liter sampling section
" network 1 " is suitable for simple scenario;The multitask network portion constitutes " network 3 ".
Specific implementation:
Step 6.1: constructing " network 2 " of complex scene, the learning difficulty of complex scene is higher, using resnet50 convolutional Neural
Network structure is trained, using the down-sampled structure of five layers of network;It rises in sampling process, in conjunction with the mentality of designing of Unet,
To each layer of progress deconvolution, it is finally reduced to the image format of 500*500*64;
Step 6.2: constructing " network 1 " of simple scenario, the network parameter that simple scenario needs is less, down-sampled and one by one
It is a to rise sampling section composition;Down-sampled part uses the structure of five layer networks, and every layer network includes two 3*3 convolutional layers and one
It is relu layers, down-sampled by max_pooling layers of progress between every layer, it is down-sampled by four layers, the image of 500*500*5 is become
For the format of 32*32*1024, a liter sampling then is carried out with two dimension deconvolution, the output result for rising sampling is 500*500*64;
Step 6.3: all liters of sampled results of step 6.1 and step 6.2 being all input to " network 3 ", are appointed for carrying out three kinds
The study respectively of business;" network 3 " mainly using 3*3 convolutional layer and relu layers of composition, input data format 500*500*64 is main
In tasking learning, output data format 500*500*5 is divided into five kinds of classifications, as a classification results;Two secondary tasks
In study, output data format 500*500*2, two secondary tasks are two classification to road and building respectively, as
Other two classification results.
Step 6.4: three classification results in step 6.3 being subjected to the calculating of loss value, main task weight is set as 1, two secondary tasks
Weight be all set as 0.4, final loss=loss1*2+loss2*0.4+loss3*0.4, and loss value is carried out
The calculating of AdamOptimizer is trained network parameter by iteration.
Three training
Step 7: by complex scene image sample obtained in step 5 and its corresponding calibration sample the input deep learning
Network, the training to the wherein second down-sampled part, the part progress parameter for rising sampling section and multitask network;Complete training
Afterwards, the training of freezing step 7 obtains model parameter, prepares for step 8 training step;
Step 8: by simple scenario image sample obtained in step 5 and its corresponding calibration sample the input deep learning
Network carries out the training of the first down-sampled network parameter, and the parameter for rising sampling and multitask trained in step 7 is frozen
Knot, the model parameter for only the first down-sampled network portion for needing to learn.
Four classification and post-processing
Step 9: carrying out the calculating of scene information entropy and variance to test section image, be divided into complex scene and simple field with threshold method
Scape, and the model parameter obtained with training carries out atural object point to the test zone high-resolution urban remote sensing image of different complexities
Class;
Step 10: it is down-sampled to the high-resolution remote sensing image progress of test section, then using canny operator to down-sampled result
Carry out the extraction of marginal information;
Step 11: the marginal information obtained according to step 10 handles the broken figure spot of the classification results of step 9, obtains
More complete terrain classification result.
2. a kind of high-resolution remote sensing image building extracting method based on multitask deep learning, which is characterized in that described
The specific implementation of step 10 is as follows:
Step 10.1: first by the high resolution image of 0.1m it is down-sampled be 0.5m, then the image after down-sampled is carried out
GaussianBlur (Gaussian smoothing) inhibits false edge;
Step 10.2: carrying out the edge detection of canny operator, obtain edge detection results.
3. a kind of high-resolution remote sensing image building extracting method based on multitask deep learning, which is characterized in that step
11 specific implementation is as follows:
Step 11.1: the classification results of step 9 being carried out with the expansion of having ready conditions of edge constraint, by the resulting edge detection of step 10
As a result it is used as exposure mask, the classification results of step 9 are expanded, if there is edge in the region in expansion process, without
Expansion calculates;
Step 11.2: classification results being carried out with the corrosion of having ready conditions of edge constraint, equally by the resulting marginal check knot of step 10
Fruit corrodes the expansion results that step 11.1 obtains as exposure mask, carries out at corrosion at pixel existing for edge for having
Reason.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910274699.3A CN109961105B (en) | 2019-04-08 | 2019-04-08 | High-resolution remote sensing image classification method based on multitask deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910274699.3A CN109961105B (en) | 2019-04-08 | 2019-04-08 | High-resolution remote sensing image classification method based on multitask deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109961105A true CN109961105A (en) | 2019-07-02 |
CN109961105B CN109961105B (en) | 2020-03-27 |
Family
ID=67025731
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910274699.3A Active CN109961105B (en) | 2019-04-08 | 2019-04-08 | High-resolution remote sensing image classification method based on multitask deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109961105B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111160127A (en) * | 2019-12-11 | 2020-05-15 | 中国资源卫星应用中心 | Remote sensing image processing and detecting method based on deep convolutional neural network model |
CN111553272A (en) * | 2020-04-28 | 2020-08-18 | 上海市测绘院 | High-resolution satellite optical remote sensing image building change detection method based on deep learning |
CN114037024A (en) * | 2022-01-07 | 2022-02-11 | 心鉴智控(深圳)科技有限公司 | Multitask neural network based data identification system and method |
CN115546649A (en) * | 2022-10-24 | 2022-12-30 | 中国矿业大学(北京) | Single-view remote sensing image height estimation and semantic segmentation multi-task prediction method |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150186792A1 (en) * | 2013-12-26 | 2015-07-02 | Xerox Corporation | System and method for multi-task learning for prediction of demand on a system |
CN105912790A (en) * | 2016-04-15 | 2016-08-31 | 重庆大学 | Depth regression model based remote sensing water quality monitoring method |
CN107833183A (en) * | 2017-11-29 | 2018-03-23 | 安徽工业大学 | A kind of satellite image based on multitask deep neural network while super-resolution and the method for coloring |
CN108052940A (en) * | 2017-12-17 | 2018-05-18 | 南京理工大学 | SAR remote sensing images waterborne target detection methods based on deep learning |
CN108550115A (en) * | 2018-04-25 | 2018-09-18 | 中国矿业大学 | A kind of image super-resolution rebuilding method |
CN109345547A (en) * | 2018-10-19 | 2019-02-15 | 天津天地伟业投资管理有限公司 | Traffic lane line detecting method and device based on deep learning multitask network |
-
2019
- 2019-04-08 CN CN201910274699.3A patent/CN109961105B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150186792A1 (en) * | 2013-12-26 | 2015-07-02 | Xerox Corporation | System and method for multi-task learning for prediction of demand on a system |
CN105912790A (en) * | 2016-04-15 | 2016-08-31 | 重庆大学 | Depth regression model based remote sensing water quality monitoring method |
CN107833183A (en) * | 2017-11-29 | 2018-03-23 | 安徽工业大学 | A kind of satellite image based on multitask deep neural network while super-resolution and the method for coloring |
CN108052940A (en) * | 2017-12-17 | 2018-05-18 | 南京理工大学 | SAR remote sensing images waterborne target detection methods based on deep learning |
CN108550115A (en) * | 2018-04-25 | 2018-09-18 | 中国矿业大学 | A kind of image super-resolution rebuilding method |
CN109345547A (en) * | 2018-10-19 | 2019-02-15 | 天津天地伟业投资管理有限公司 | Traffic lane line detecting method and device based on deep learning multitask network |
Non-Patent Citations (2)
Title |
---|
刘大伟等: ""基于深度学习的高分辨率遥感影像分类研究"", 《光学学报》 * |
杨诗寓: ""基于深度学习的多源遥感影像目标检测技术的研究"", 《中国优秀硕士学位论文全文数据库 工程科技II辑》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111160127A (en) * | 2019-12-11 | 2020-05-15 | 中国资源卫星应用中心 | Remote sensing image processing and detecting method based on deep convolutional neural network model |
CN111160127B (en) * | 2019-12-11 | 2023-07-21 | 中国四维测绘技术有限公司 | Remote sensing image processing and detecting method based on deep convolutional neural network model |
CN111553272A (en) * | 2020-04-28 | 2020-08-18 | 上海市测绘院 | High-resolution satellite optical remote sensing image building change detection method based on deep learning |
CN111553272B (en) * | 2020-04-28 | 2022-05-06 | 上海市测绘院 | High-resolution satellite optical remote sensing image building change detection method based on deep learning |
CN114037024A (en) * | 2022-01-07 | 2022-02-11 | 心鉴智控(深圳)科技有限公司 | Multitask neural network based data identification system and method |
CN115546649A (en) * | 2022-10-24 | 2022-12-30 | 中国矿业大学(北京) | Single-view remote sensing image height estimation and semantic segmentation multi-task prediction method |
Also Published As
Publication number | Publication date |
---|---|
CN109961105B (en) | 2020-03-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111986099B (en) | Tillage monitoring method and system based on convolutional neural network with residual error correction fused | |
CN109961105A (en) | A kind of Classification of High Resolution Satellite Images method based on multitask deep learning | |
CN109446992B (en) | Remote sensing image building extraction method and system based on deep learning, storage medium and electronic equipment | |
Chen et al. | Road extraction in remote sensing data: A survey | |
CN112287807B (en) | Remote sensing image road extraction method based on multi-branch pyramid neural network | |
CN107067405B (en) | Remote sensing image segmentation method based on scale optimization | |
CN112016436A (en) | Remote sensing image change detection method based on deep learning | |
CN106529508A (en) | Local and non-local multi-feature semantics-based hyperspectral image classification method | |
CN115170979B (en) | Mining area fine land classification method based on multi-source data fusion | |
CN107832797B (en) | Multispectral image classification method based on depth fusion residual error network | |
CN105139395A (en) | SAR image segmentation method based on wavelet pooling convolutional neural networks | |
CN113312993B (en) | Remote sensing data land cover classification method based on PSPNet | |
CN110956207B (en) | Method for detecting full-element change of optical remote sensing image | |
CN110991430B (en) | Ground feature identification and coverage rate calculation method and system based on remote sensing image | |
CN112767413B (en) | Remote sensing image depth semantic segmentation method integrating region communication and symbiotic knowledge constraints | |
CN113919226B (en) | Mining vegetation ecological cumulative effect disturbance range identification method based on weight | |
CN111738052A (en) | Multi-feature fusion hyperspectral remote sensing ground object classification method based on deep learning | |
CN112037244B (en) | Landsat-8 image culture pond extraction method combining index and contour indicator SLIC | |
CN114037891A (en) | High-resolution remote sensing image building extraction method and device based on U-shaped attention control network | |
CN111553272B (en) | High-resolution satellite optical remote sensing image building change detection method based on deep learning | |
Shackelford et al. | Urban road network extraction from high-resolution multispectral data | |
CN114913430A (en) | Object height identification method based on high resolution image and Unet shadow detection | |
CN113515969B (en) | Multi-temporal remote sensing image earth surface coverage classification method based on convolutional neural network | |
CN114596490A (en) | Hilly land feature line extraction method and hilly land DEM (digital elevation model) fine production method | |
CN113963265A (en) | Small sample small target rapid detection and identification method for complex remote sensing land environment |
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 |