CN109784209A - Utilize the high and cold mountain area accumulated snow extracting method of high-resolution remote sensing image - Google Patents
Utilize the high and cold mountain area accumulated snow extracting method of high-resolution remote sensing image Download PDFInfo
- Publication number
- CN109784209A CN109784209A CN201811602060.5A CN201811602060A CN109784209A CN 109784209 A CN109784209 A CN 109784209A CN 201811602060 A CN201811602060 A CN 201811602060A CN 109784209 A CN109784209 A CN 109784209A
- Authority
- CN
- China
- Prior art keywords
- accumulated snow
- layer
- remote sensing
- sensing image
- image
- 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
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to remote sensing technology fields, disclose a kind of high and cold mountain area accumulated snow extracting method using high-resolution remote sensing image, the following steps are included: step S1), high-resolution remote sensing image is split using the fractal net work evolution partitioning algorithm of optimal scale, obtains high resolution image object;Step S2), higher-dimension imaged object characteristic pattern is constructed using high resolution image object;Step S3), construct depth confidence network model;Step S4), construct full condition of contact random field models;Step S5), output accumulated snow extracts result.Method of the invention is stronger to the applicability of different sensors data, and the accumulated snow for being more suitable for high resolution remote sensing image extracts;Depth confidence network is combined with full condition of contact random field models using the thought of object-oriented, makes full use of spectrum, shape, texture and the spatial relationship information of high-resolution image, the spiced salt phenomenon in accumulated snow extraction result is eliminated, improves the precision that accumulated snow extracts.
Description
Technical field
The present invention relates to remote sensing technology fields, and in particular to a kind of high and cold mountain area accumulated snow using high-resolution remote sensing image
Extracting method.
Background technique
It is restricted by factors such as topographic and geologic and environmental protections, the more difficult quantitative acquisition of method of traditional geography, meteorology
The accumulated snow data of high and cold mountain area.With the fast development of remote sensing technology, new technical means are provided for Monitoring Snow Cover.
Current MODIS data and its snow lid product are widely used in Monitoring Snow Cover." is sentenced using MODIS data and know Qilian
Mountain area accumulated snow technique study " benefit is disclosed in (Wang Xing, Zhang Qiang, Guo's niobium etc., " Arid Meteorology " in by the end of June, 2007 the phase of volume 25 the 2nd)
With the MODIS image of Qilian Mountain Area, it is based on normalization difference snow index (Normalized Difference Snow
Index, NDSI) exponent extracting accumulated snow region." resolution sub-pixed mapping avenges charting algorithm research in Qinghai-Tibet Platean " (Zhang Hongen, in
Remote sensing application research institute, graduate school, the academy of sciences, state doctoral thesis in 2004) in disclose based on MODIS data using mixing point
Solution is theoretical, develops automatically snow charting algorithm." the Snow Cover Area inverting research based on NDSI-NDVI feature space " (Chen Wenqian,
Ding Jianli, Sun Yong are violent etc., " dirt band " in August, 2015 the 4th phase of volume 38) in disclose normalized site attenuation
(Normalized Difference Vegetation Index, NDVI) is combined with NDSI index, establishes accumulated snow inverting mould
Type, to improve accumulated snow extraction accuracy.
In view of the highest spatial resolution of MODIS data only has 500m, it is more difficult to be applicable in fining Monitoring Snow Cover, but by
Short infrared wave band needed for lacking classical snow index NDSI in high resolution image, therefore tradition is mentioned based on NDSI index
It takes the method for accumulated snow and is not suitable for.
Develop the substitution index of NDSI to carry out accumulated snow to extract being a kind of common way."Detection of
spatial,temporal,and spectral surface changes in the Ny-area 79°N,
Svalbard,using a low cost multispectral camera in combination with
spectroradiometer measurements》(Hinkler J,J B, Hansen B U., " Physics&
Chemistry of the Earth Parts A/b/c " the 28th phase in 2003) in propose RGBNDSI and NDSII index generation
Snow detection is carried out for NDSI index." GF-1 satellite remote sensing snow face reflectivity calculates under the MODEL OVER COMPLEX TOPOGRAPHY of mountain area " (Jiang Luyuan,
Xiao Pengfeng, Feng Xuezhi etc., " Nanjing University's journal (natural science) " in September, 2015 the 5th phase of volume 51) in propose in NDSII
On the basis of index, building is suitable for the accumulated snow extracting index of GF-1 satellite, and experiment shows that the index can effectively identify accumulated snow.
However influenced by mountainous shade, such method often need to carry out topographical correction using high accuracy DEM, could effectively identify mountain area product
Snow.
The method of image classification is to realize the another kind of method of high-resolution remote sensing image snow detection."Taylor&
Francis Online::Automatic snow cover monitoring at high temporal and spatial
resolution,using images taken by a standard digital camera-International
Journal of Remote Sensing " (J.Hinkler, S.B.Pedersen, M.Rasch, " International
Journal of Remote Sensing " the 21st phase of volume 23 in 2012) in disclose using IKONOS image analysing computer accumulated snow with
The SPECTRAL DIVERSITY of other atural object classifications, and then accumulated snow information is extracted using ISODATA algorithm."Support vector
machine-based decision tree for snow cover extraction in mountain areas using
High spatial resolution remote sensing image " (Zhu L., " Journal of Applied
Remote Sensing " the 1st phase of volume 8 in 2014) in mention based on the multiple snow detection features of ZY-3 Extraction of Image, using certainly
Plan tree method identifies Northern Piedmont of Tianshan Mountains accumulated snow." the satellite image Extraction of snow information of object-oriented " (Yong Wanling, poplar text, Zhang Li
Peak etc., " Surveying and mapping " the 9th phase of volume 41 in 2016) in disclose using HJ image using Object--oriented method extract accumulated snow
Information.
From the foregoing, it will be observed that being limited by the shortage of short infrared wave band and high and cold mountain area terrain data, the side of image classification is utilized
Method is to extract the effective ways of the high and cold mountain area accumulated snow information of high resolution image.Furthermore in view of pixel-based using tradition
Easily there is spiced salt phenomenon in the accumulated snow region that method extracts on high-resolution image, and the texture of high-resolution image, geometry and space are closed
The abundant informations such as system, it is urgent to provide a kind of high-resolution image high and cold mountain area accumulated snow extracting methods to extract result to effectively eliminate accumulated snow
In spiced salt phenomenon, to obtain accurate continuous accumulated snow area.
Summary of the invention
The object of the present invention is to provide a kind of high and cold mountain area accumulated snow extracting method using high-resolution remote sensing image,
It solves to effectively eliminate the spiced salt phenomenon in accumulated snow extraction result, obtains accurate continuous accumulated snow area.
To achieve the above object, the high and cold mountain area accumulated snow extraction side using high-resolution remote sensing image designed by the present invention
Method, comprising the following steps: step S1), using the fractal net work evolution partitioning algorithm of optimal scale to high-resolution remote sensing image
It is split, obtains high resolution image object;Step S2), it is special using high resolution image object building higher-dimension imaged object
Sign figure;Step S3), construct depth confidence network model;Step S4), construct full condition of contact random field models;Step S5), it is defeated
Accumulated snow extracts result out.
Preferably, the step S1) step 1.1) is further included steps of, from accumulated snow to be extracted
Remote sensing image in cut out test block for determining optimum segmentation scale;Step 1.2) sets gradually segmentation scale, to reality
It tests area's image and carries out fractal net work evolution segmentation, the wave band average for calculating every layer of each object is poor;Step 1.3), statistics are each
The wave band average of layer object is poor, calculates the change rate of standard deviation under different scale layer;Step 1.4) takes change rate with scale
The peak of curve of variation is optimum segmentation scale.
Preferably, in the step 1.1), size is cut out from the remote sensing image of accumulated snow to be extracted
For the accumulated snow region of 1000 pixel *, 1000 pixel, the test block as the determining image optimum segmentation scale;In the step
1.2) in, with step-length 3 for scale interval, corresponding segmentation scale is successively set from 0 to 200, FNEA is carried out to test block image
Multi-scale division counts the number of every layer of cutting object and the wave band standard deviation of each object, and the wave band for calculating every layer of object is flat
Equal standard deviation SD.
Preferably, in the step 1.3), the value of the wave band average difference SD of each layer object is counted,
The change rate ROC of the SD under different scale layer is calculated according to following formula:
Wherein, i is object level number, SD(i)Average for current layer object is poor, SD(i-1)For being averaged for next layer of object
Objective metric is poor.
Preferably, in the step 1.5), optimum segmentation scale is extracted using determining accumulated snow, is used
FNEA algorithm carries out multi-scale division to whole picture remote sensing image and obtains segmentation result.
Preferably, in the step 2), spectral signature, shape feature and the texture of imaged object are extracted
Feature.
Preferably, the step S3) step 3.1) is further included steps of, in accumulated snow to be extracted
High-resolution remote sensing image on choose ground class sample as training sample train obtain bottom RBM model;Step 3.2), benefit
The hidden layer for the bottom RBM model for using step 3.1) to obtain is as the 1st layer of visible layer, and using contrast divergence algorithm training
Obtain the 1st layer of RBM model;Step 3.3), using the hidden layer of (i-1)-th layer of RBM model as i-th layer of visible layer, using pair
The training of sdpecific dispersion algorithm obtains i-th layer of RBM model, until current layer number i reaches number of plies n set by DBN model;Step
3.4), optimize DBN model parameter;Step 3.5), the DBN model after obtaining parameter optimization.
Preferably, in the step 3.4), a BP neural network is added in the top layer of DBN model
Layer, BP neural network layer is trained using training sample, by successively backward in the way of propagate every layer training generate mistake
Difference, and the parameter of DBN is finely adjusted using stochastic gradient descent method;In the step 3.5), BP neural network layer is repeated
Training is until error reaches setting value or reaches maximum cycle, the DBN model after obtaining parameter optimization.
Preferably, in the step S4) in, utilize the step S3) in the obtained object of DBN model
Class probability defines single order potential function Ψi(Oi):
Ψi(Oi)=- In (PDBN(Oi=k)),
In formula, PDBN(Oi=k) be the step S3) in DBN model output i-th of object OiProbability when classification is k;
Second order potential function Ψ is defined using two different gaussian kernel function linear combinationsij(Oi,Oj)
Ψij(Oi,Oj)=μ (Oi,Oj)(ω1f1(Oi,Oj)+ω2f2(Oi,Oj)),
Wherein,
In formula, μ (Oi,Oj) it is classification compatibility weight;ω1With ω2For Gauss weight coefficient;piAnd pjDivide than being object Oi
And OjSpatial position in the picture, with the mean value calculation of all pixels spatial position in object;θiAnd θjFor object OiWith
OjAll wave bands composition feature vector, obtained with the mean value calculation of the wave band feature vector of all pixels in object;α,
β, γ are used to regulating object OiAnd OjPosition and wave band feature vector relative size;
Above-mentioned parameter is obtained using the method for segmental training using selected sample, and then obtains the full connection in conjunction with DBN
Condition random field.
Preferably, in the step S5) in, using mean field algorithm to the step S4) in construct
Inferred in conjunction with the full condition of contact random field models of DBN model, calculate the class probability of each object, and according to maximum after
It tests criterion and judges whether each object type is accumulated snow area.
The beneficial effects of the present invention are: the high and cold mountain area accumulated snow extracting method of the invention using high-resolution remote sensing image
It is stronger to the applicability of different sensors data, and the accumulated snow for being more suitable for high resolution remote sensing image extracts;Utilize object-oriented
Thought depth confidence network is combined with full condition of contact random field models, make full use of spectrum, the shape of high-resolution image
Shape, texture and spatial relationship information eliminate the spiced salt phenomenon in accumulated snow extraction result, improve the precision that accumulated snow extracts.
Detailed description of the invention
Fig. 1 is the stream of the high and cold mountain area accumulated snow extracting method using high-resolution remote sensing image of the preferred embodiment of the present invention
Cheng Tu.
Fig. 2 is the sub-process figure of the optimal scale fractal net work evolution segmentation high-resolution remote sensing image step in Fig. 1.
Fig. 3 is the structural schematic diagram of depth confidence network model.
Fig. 4 is the training schematic diagram of depth confidence network.
Fig. 5 is the sub-process figure of the building depth confidence network model step in Fig. 1.
Specific embodiment
The following further describes the present invention in detail with reference to the accompanying drawings and specific embodiments.
Lack short infrared wave band needed for calculating NDSI index for high resolution remote sensing image, is extracted to eliminate accumulated snow
As a result the spiced salt phenomenon in, the bright high and cold mountain area accumulated snow extracting method using high-resolution remote sensing image of we are used towards right
The combination depth confidence network (Deep Belief Network, DBN) of elephant and full condition of contact random field (Conditional
Random Fields, CRF) high-resolution image high and cold mountain area accumulated snow extracting method extract knot to obtain accurate continuous accumulated snow
Fruit, as shown in fig. 1, comprising the following steps:
Step S1), the optimal scale fractal net work of high-resolution remote sensing image, which develops, to be divided.For accumulated snow information to be extracted
High-resolution remote sensing image, it is first determined the optimum segmentation scale of image, then using fractal net work evolution algorithmic generate it is distant
Imaged object is felt, as accumulated snow extraction unit.
Step S2), construct higher-dimension imaged object characteristic pattern.Using the imaged object of generation as unit, the light of each object is calculated
Spectrum, shape, Texture eigenvalue make full use of the characteristic information of high resolution image, and then construct higher-dimension imaged object characteristic pattern
Picture.
Step S3), construct depth confidence network model.The number of plies of DBN model, each node layer number are set, chosen of all categories
Sample is based on higher-dimension imaged object characteristic pattern, training DBN model.
Step S4), construct full condition of contact random field models.Utilize the class for each object that trained DBN model obtains
The single order potential function of other probability building condition random field, utilizes the second-order potential of gaussian kernel function model construction conditional random field models
Function, and the conditional random field models for obtaining model relevant parameter by segmental training, and then being connected entirely.
Step S5), output accumulated snow extracts result.It is pushed away using conditional random field models of the mean field algorithm to building
It is disconnected, the class probability of each object is calculated, and judge each object type according to maximum a posteriori criterion, the accumulated snow finally extracted
Area.
Below in conjunction with attached drawing to the high and cold mountain area accumulated snow using high-resolution remote sensing image of the preferred embodiment of the present invention
Each step of extracting method will carry out further parsing in detail respectively.
Step S1), the optimal scale fractal net work of high resolution remote sensing image, which develops, to be divided.
Divide (Fractal Network Evolution Algorithm, FNEA) algorithm in view of fractal net work develops
It is a kind of multi-scale division algorithm, and spectrum, shape and the texture information of image can be made full use of, the present invention utilizes the algorithm structure
Build high-resolution imaged object.The basic thought of FNEA be by fractal iteration process, according to the maximum merging criterion of similitude, from
Pixel starts two adjacent objects for merging the condition that meets, and key problem in technology is the definition of the similarity criterion between adjacent object.
The similitude of object is integrated spectral similitude and shape similarity to measure in classical FNEA, wherein the similarity measurements of spectrum
Measure hvalIt is to be measured using the standard deviation of grey scale pixel value in object, the similitude h of shapeshaUsing shape feature before and after merging
Variation in space describes.Then the integrated spectral of adjacent two object and the similarity criterion of shape may be defined as:
F=wshahsha+(1-wsha)hval
In above formula, wshaFor weight, f is the segmentation scale in FNEA algorithm, and giving different segmentation scale threshold values can obtain
To different size of cutting object, the object number that the bigger segmentation of threshold value obtains is smaller, and object is also bigger.
The size of segmentation scale directly affects the quality of FNEA algorithm segmentation result, and segmentation result directly affect it is subsequent
The precision that accumulated snow extracts.To determine the optimum segmentation scale for extracting accumulated snow, sentenced using the average difference of every layer of cutting object
The superiority and inferiority of other segmentation result, and then the optimum segmentation scale that accumulated snow extracts is obtained, finally generated using fractal net work evolution algorithmic
Remote sensing image object, as subsequent accumulated snow extraction unit.The detailed process of the optimal scale segmentation of High spatial resolution remote sensing is as follows
It is shown:
Step 1.1) cuts out the test block of determining optimum segmentation scale from the remote sensing image of accumulated snow to be extracted.Because distant
The optimum segmentation scale for feeling different land types on image is generally different, for the optimum segmentation ruler for preferably determining image accumulated snow extraction
Degree cuts out the accumulated snow region that size is 1000 pixel *, 1000 pixel from the remote sensing image of accumulated snow to be extracted, is used as determining be somebody's turn to do
The test block of image optimum segmentation scale.
Step 1.2) sets gradually segmentation scale and carries out fractal net work evolution segmentation to test block image, and every layer of calculating is right
The wave band average of elephant is poor.With step-length 3 for scale interval, corresponding segmentation scale is successively set from 0 to 200, to test block
Image carries out FNEA multi-scale division, counts the number of every layer of cutting object and the wave band standard deviation (Standard of each object
Deviation, SD), and calculate the wave band average difference SD of every layer of object.
Step 1.3), the wave band average for counting each layer object is poor, calculates the change rate of standard deviation under different scale layer.
The value of the wave band average difference SD of each layer object is counted, the change rate ROC of the SD under different scale layer is calculated:
Wherein, i is object level number, i.e. SD(i)Average for current layer object is poor, SD(i-1)For the flat of next layer of object
Equal objective metric is poor.
Step 1.4), take change rate with dimensional variation peak of curve be optimum segmentation scale.Change rate ROC is drawn with ruler
Spend the curve graph of variation, when peak value, that is, ROC value maximum of curve corresponding scale-value, the most optimal sorting extracted for the image accumulated snow
Cut scale.
Step 1.5) carries out multi-scale division to whole picture remote sensing image using optimum segmentation scale.It is true using the above method
Fixed accumulated snow extracts optimum segmentation scale, carries out multi-scale division to whole picture remote sensing image using FNEA algorithm, obtains final
Segmentation result.
Step S2), construct higher-dimension imaged object characteristic pattern.
Different types of high resolution sensor load variations are larger, and high resolution image wave band is less, general to only have
Visible light wave range, near infrared band and panchromatic wave-band;In addition the accumulated snow spectral characterization on the remote sensing image under different terrain conditions
It is general different, also easily occur " the different spectrum of jljl " and " same object different images " phenomenon on image, single utilization accumulated snow spectral signature is more difficult straight
Connect the foundation as high resolution remote sensing image snow detection.Compared with single pixel, High Resolution Satellite Images object includes abundant
Characteristic information make full use of each characteristic information of imaged object such as the geological information, texture information and spectral information of object
Help preferably to identify accumulated snow.For this purpose, imaged object is generated using FNEA multi-scale division algorithm obtained in step S1),
And extract spectrum, shape and the textural characteristics of imaged object.Feature used in the method for the present invention is as shown in table 1:
Object image feature used in 1 the method for the present invention of table
Step S3), construct depth confidence network model.
Depth confidence network (Deep Belief Network, DBN) is suitble to high dimensional feature information and complicated classification problem
Feature extraction and modeling are carried out, and has been successfully applied to remote sensing image classification field.DBN is by the restricted Boltzmann of multilayer
Machine (Restricted Boltzmann Machine, RBM) stacks composition, and wherein RBM model is by visible layer and hidden layer institute
Two layers of the neural network constituted.DBN is using training sample as the input of bottom RBM visible layer, and after training sample is learnt
Input of the hidden layer of acquisition as one layer of RBM visible layer thereon, and then the hidden layer of this layer of RBM is obtained, then this is implied
Input of the layer as one layer of RBM visible layer thereon;And so on, the DBN network mould of multilayer is obtained by the method for Level by level learning
Type.For DBN model is applied in Classification in Remote Sensing Image, the top layer of DBN model add BP (Back Propagation,
Backpropagation) layer, and then the output of top layer RBM model is exported by BP layers as a result, this hair as BP layers of input
The network structure of the bright DBN model utilized is as shown in Figure 3.
The training process of above-mentioned DBN model is divided into two processes of pre-training and fine tuning, and training process is as shown in Figure 4.First
By individually training to each layer of RBM model, then the preliminary parameters for obtaining DBN network model are carried out using sample
Supervised learning optimization fine tuning DBN network model parameter, and then complete the training to entire DBN network model.Please refer to figure
5, the present invention constructs depth confidence network model, and specific step is as follows:
Step 3.1) chooses ground class sample as training sample training on the high-resolution remote sensing image of accumulated snow to be extracted
Obtain bottom RBM model.On the remote sensing image of accumulated snow to be extracted, the different types of sample of each ground class, every kind of ground are chosen
Class preferably chooses multiple pure samples.Using the visible layer of the training sample initialization bottom RBM of selection, and dissipated using comparison
Degree algorithm training obtains bottom RBM model.
Step 3.2), the hidden layer of the bottom RBM model obtained using step 3.1) as the 1st layer of visible layer, and
1st layer of RBM model is obtained using contrast divergence algorithm training.
Step 3.3) is instructed using the hidden layer of (i-1)-th layer of RBM model as i-th layer of visible layer using contrast divergence algorithm
The RBM model to i-th layer is got, until current layer number i reaches number of plies n set by DBN model.
Step 3.4) optimizes DBN model parameter.Last BP neural network layer is trained using training sample, is pressed
The error that every layer of training generates is propagated according to mode successively backward, and micro- using parameter progress of the stochastic gradient descent method to DBN
It adjusts.
Step 3.5), the DBN model after obtaining parameter optimization.The training of BP neural network layer is repeated, is set until error reaches
Definite value reaches maximum cycle, the DBN model after obtaining parameter optimization at this time.
Step S4), construct full condition of contact random field models.
Condition random field (Conditional Random Fields, CRF) model is a kind of random field models of classics,
Because it has unique advantage in any multiple features fusion and spatial context information representation, it is widely used in the segmentation of image, divides
The fields such as class, identification.But traditional CRF method causes ground class boundary information it is easier that smooth phenomenon occurred in classification results
It loses.For this purpose, the present invention is taxon using the object that segmentation in above-mentioned steps S1) generates, CRF mould is utilized on this basis
Type carries out accumulated snow extraction.
CRF is a kind of discriminate probability graph model, generallys use the neighborhood phase interaction that Gibbs distribution comes between analog variable
With.It is to find to make the maximum image category label of posterior probability using the basic thought that CRF frame is classified, further can be exchanged into
Calculate image category label when corresponding energy function being made to reach minimum value.The present invention is using object as basic observing unit, then
The expression formula of corresponding energy function is as follows:
Ψ in formulai(Oi) it is single order potential function, Ψij(Oi,Oj) it is second order potential function, N is the object number of image, Oi、Oj
Respectively i-th and j object of image,For object OiNeighborhood object label.
From the above equation, we can see that the building of potential function is the key that CRF model, wherein single order potential function is usually for single-point
The corresponding class label modeling of observation information, the present invention utilizes above-mentioned steps S3 thus) in the obtained object of DBN model
Class probability is to define single order potential function Ψi(Oi), expression formula is as follows:
Ψi(Oi)=- In (PDBN(Oi=k))
In formula, PDBN(Oi=k) be step S3) in DBN model output i-th of object OiProbability when classification is k.
Spatial context information of the second order potential function commonly used to description image, common way are using Potts model
It defines, but this method takes into consideration only the neighborhood object of observation object, and underuses the spatial relationship between object.For this purpose,
The present invention defines second order potential function Ψ using two different gaussian kernel function linear combinationsij(Oi,Oj), connect entirely to constitute
The condition random field connect, expression formula are as follows:
Ψij(Oi,Oj)=μ (Oi,Oj)(ω1f1(Oi,Oj)+ω2f2(Oi,Oj))
Wherein:
In above-mentioned expression formula, μ (Oi, Oj) it is classification compatibility weight;ω1With ω2For Gauss weight coefficient;piAnd pjRespectively
For object OiAnd OjSpatial position in the picture, with the mean value calculation of all pixels spatial position in object;θiAnd θjFor
Object OiAnd OjAll wave bands composition feature vector, with the mean value calculation of the wave band feature vector of all pixels in object
It obtains, α, β, γ are used to regulating object OiAnd OjPosition and wave band feature vector relative size;It is used using selected sample
Above-mentioned parameter can be obtained in the method for segmental training, and then obtains the full condition of contact random field of combination DBN constructed by this method.
Step S5), output category result.
It is carried out using full condition of contact random field models of the mean field algorithm to the combination DBN model constructed in step S4)
Infer, calculates the class probability of each object, and judge whether each object type is accumulated snow area according to maximum a posteriori criterion, in turn
Obtain the accumulated snow area of whole picture image.Specific step is as follows:
Step 5.1) obtains the classification of each object.The characteristic vector for inputting each object, using mean field algorithm to condition with
Airport model is inferred, the class probability of each object is successively obtained, and according to maximum posteriori criterion, obtains the object
Classification.
Step 5.2) obtains the classification information of whole picture image.Using object as taxon, pixel each in object is marked
For same classification;It repeats the above steps, and then obtains the classification information of whole picture image, and exporting is single-range classification shadow
Picture.
Step 5.3) obtains the distribution of Snow Cover Over figure of image.Using the drawing function of GIS software, classification image is carried out special
Drawing designing finally obtains the distribution of Snow Cover Over figure of the image.
Compared with prior art, the high and cold mountain area accumulated snow extracting method of the invention using high-resolution remote sensing image has
It has the advantage that
(1) accumulated snow is extracted using the method for NDSI snow index compared to tradition, this method fits different sensors data
It is stronger with property, and the accumulated snow for being more suitable for high resolution remote sensing image extracts;
(2) the accumulated snow extracting method of image classification is utilized compared to tradition, this method utilizes the thought of object-oriented by depth
Confidence network is combined with full condition of contact random field models, makes full use of spectrum, shape, texture and the space of high-resolution image
Relation information eliminates the spiced salt phenomenon in accumulated snow extraction result, improves the precision that accumulated snow extracts.
The embodiments described above only express several embodiments of the present invention, and the description thereof is more specific and detailed, but simultaneously
Limitations on the scope of the patent of the present invention therefore cannot be interpreted as.It should be pointed out that for those of ordinary skill in the art
For, without departing from the inventive concept of the premise, various modifications and improvements can be made, these belong to guarantor of the invention
Protect range.Therefore, the scope of protection of the patent of the invention shall be subject to the appended claims.
Claims (10)
1. a kind of high and cold mountain area accumulated snow extracting method using high-resolution remote sensing image, comprising the following steps:
Step S1), high-resolution remote sensing image is split using the fractal net work evolution partitioning algorithm of optimal scale, is obtained
High resolution image object;
Step S2), higher-dimension imaged object characteristic pattern is constructed using high resolution image object;
Step S3), construct depth confidence network model;
Step S4), construct full condition of contact random field models;
Step S5), output accumulated snow extracts result.
2. the high and cold mountain area accumulated snow extracting method according to claim 1 using high-resolution remote sensing image, feature exist
It is further included steps of in the step S1)
Step 1.1) cuts out the test block for determining optimum segmentation scale from the remote sensing image of accumulated snow to be extracted;
Step 1.2) sets gradually segmentation scale, carries out fractal net work evolution segmentation to test block image, every layer of calculating is each right
The wave band average of elephant is poor;
Step 1.3), the wave band average for counting each layer object is poor, calculates the change rate of standard deviation under different scale layer;
Step 1.4), take change rate with dimensional variation peak of curve be optimum segmentation scale.
3. the high and cold mountain area accumulated snow extracting method according to claim 2 using high-resolution remote sensing image, feature exist
In: in the step 1.1), the product that size is 1000 pixel *, 1000 pixel is cut out from the remote sensing image of accumulated snow to be extracted
Region is avenged, as the test block for determining the image optimum segmentation scale;In the step 1.2), with step-length 3 for scale interval,
Corresponding segmentation scale is successively set from 0 to 200, FNEA multi-scale division is carried out to test block image, counts every layer of segmentation pair
The wave band standard deviation of the number of elephant and each object, and calculate the wave band average difference SD of every layer of object.
4. the high and cold mountain area accumulated snow extracting method according to claim 2 using high-resolution remote sensing image, feature exist
In: in the step 1.3), the value of the wave band average difference SD of each layer object is counted, calculates different rulers according to following formula
Spend the change rate ROC of the SD under layer:
Wherein, i is object level number, SD(i)Average for current layer object is poor, SD(i-1)For the mean object of next layer of object
Standard deviation.
5. the high and cold mountain area accumulated snow extracting method according to claim 2 using high-resolution remote sensing image, feature exist
In: in the step 1.5), extract optimum segmentation scale using determining accumulated snow, with FNEA algorithm to whole picture remote sensing image into
Row multi-scale division obtains segmentation result.
6. the high and cold mountain area accumulated snow extracting method according to claim 2 using high-resolution remote sensing image, feature exist
In: in the step 2), extract spectral signature, shape feature and the textural characteristics of imaged object.
7. the high and cold mountain area accumulated snow extracting method according to claim 2 using high-resolution remote sensing image, feature exist
It is further included steps of in the step S3)
Step 3.1) is chosen ground class sample on the high-resolution remote sensing image of accumulated snow to be extracted and is obtained as training sample training
Bottom RBM model;
Step 3.2), the hidden layer of the bottom RBM model obtained using step 3.1) are used as the 1st layer of visible layer
Contrast divergence algorithm training obtains the 1st layer of RBM model;
Step 3.3), using the hidden layer of (i-1)-th layer of RBM model as i-th layer of visible layer, trained using contrast divergence algorithm
To i-th layer of RBM model, until current layer number i reaches number of plies n set by DBN model;
Step 3.4) optimizes DBN model parameter;
Step 3.5), the DBN model after obtaining parameter optimization.
8. the high and cold mountain area accumulated snow extracting method according to claim 7 using high-resolution remote sensing image, feature exist
In: in the step 3.4), a BP neural network layer is added in the top layer of DBN model, using training sample to BP mind
Be trained through network layer, by successively backward in the way of propagate every layer training generate error, and use stochastic gradient descent
Method is finely adjusted the parameter of DBN;In the step 3.5), the training of BP neural network layer is repeated until error reaches setting value
Or reach maximum cycle, the DBN model after obtaining parameter optimization.
9. the high and cold mountain area accumulated snow extracting method according to claim 7 using high-resolution remote sensing image, feature exist
In: in the step S4) in, utilize the step S3) in DBN model obtained object type probability define single order gesture letter
Number Ψi(Oi):
Ψi(Oi)=- In (PDBN(Oi=k)),
In formula, PDBN(Oi=k) be the step S3) in DBN model output i-th of object OiProbability when classification is k;
Second order potential function Ψ is defined using two different gaussian kernel function linear combinationsij(Oi,Oj):
Ψij(Oi,Oj)=μ (Oi,Oj)(ω1f1(Oi,Oj)+ω2f2(Oi,Oj)),
Wherein,
In formula, μ (Oi,Oj) it is classification compatibility weight;ω1With ω2For Gauss weight coefficient;piAnd pjDivide than being object OiAnd Oj
Spatial position in the picture, with the mean value calculation of all pixels spatial position in object;θiAnd θjFor object OiAnd Oj's
The feature vector of all wave band compositions, is obtained with the mean value calculation of the wave band feature vector of all pixels in object;α,β,γ
For regulating object OiAnd OjPosition and wave band feature vector relative size;
Above-mentioned parameter is obtained using the method for segmental training using selected sample, and then obtains the full condition of contact in conjunction with DBN
Random field.
10. the high and cold mountain area accumulated snow extracting method according to claim 9 using high-resolution remote sensing image, feature exist
In: in the step S5) in, using mean field algorithm to the step S4) in construct combination DBN model full condition of contact
Random field models are inferred, calculate the class probability of each object, and judge that each object type is according to maximum a posteriori criterion
No is accumulated snow area.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811602060.5A CN109784209B (en) | 2018-12-26 | 2018-12-26 | Method for extracting accumulated snow in alpine mountain area by using high-resolution remote sensing image |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811602060.5A CN109784209B (en) | 2018-12-26 | 2018-12-26 | Method for extracting accumulated snow in alpine mountain area by using high-resolution remote sensing image |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109784209A true CN109784209A (en) | 2019-05-21 |
CN109784209B CN109784209B (en) | 2021-06-01 |
Family
ID=66497694
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811602060.5A Active CN109784209B (en) | 2018-12-26 | 2018-12-26 | Method for extracting accumulated snow in alpine mountain area by using high-resolution remote sensing image |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109784209B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111325771A (en) * | 2020-02-17 | 2020-06-23 | 武汉大学 | High-resolution remote sensing image change detection method based on image fusion framework |
CN114821361A (en) * | 2022-06-24 | 2022-07-29 | 成都信息工程大学 | Method and device for calculating snow depth, computer equipment and readable storage medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101799561A (en) * | 2010-02-05 | 2010-08-11 | 民政部国家减灾中心 | Snow disaster remote sensing monitoring simulation evaluation method based on disaster reduction small satellite |
CN106951909A (en) * | 2016-11-16 | 2017-07-14 | 中国科学院遥感与数字地球研究所 | A kind of snow detection method of the satellite remote-sensing images of GF 4 |
CN106991411A (en) * | 2017-04-17 | 2017-07-28 | 中国科学院电子学研究所 | Remote Sensing Target based on depth shape priori becomes more meticulous extracting method |
CN108062756A (en) * | 2018-01-29 | 2018-05-22 | 重庆理工大学 | Image, semantic dividing method based on the full convolutional network of depth and condition random field |
-
2018
- 2018-12-26 CN CN201811602060.5A patent/CN109784209B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101799561A (en) * | 2010-02-05 | 2010-08-11 | 民政部国家减灾中心 | Snow disaster remote sensing monitoring simulation evaluation method based on disaster reduction small satellite |
CN106951909A (en) * | 2016-11-16 | 2017-07-14 | 中国科学院遥感与数字地球研究所 | A kind of snow detection method of the satellite remote-sensing images of GF 4 |
CN106991411A (en) * | 2017-04-17 | 2017-07-28 | 中国科学院电子学研究所 | Remote Sensing Target based on depth shape priori becomes more meticulous extracting method |
CN108062756A (en) * | 2018-01-29 | 2018-05-22 | 重庆理工大学 | Image, semantic dividing method based on the full convolutional network of depth and condition random field |
Non-Patent Citations (1)
Title |
---|
邓海龙: "高分辨率遥感图像面向对象分割与分类方法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111325771A (en) * | 2020-02-17 | 2020-06-23 | 武汉大学 | High-resolution remote sensing image change detection method based on image fusion framework |
CN111325771B (en) * | 2020-02-17 | 2022-02-01 | 武汉大学 | High-resolution remote sensing image change detection method based on image fusion framework |
CN114821361A (en) * | 2022-06-24 | 2022-07-29 | 成都信息工程大学 | Method and device for calculating snow depth, computer equipment and readable storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN109784209B (en) | 2021-06-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Balaguer et al. | Definition of a comprehensive set of texture semivariogram features and their evaluation for object-oriented image classification | |
CN104751478B (en) | Object-oriented building change detection method based on multi-feature fusion | |
CN102646200B (en) | Image classifying method and system for self-adaption weight fusion of multiple classifiers | |
CN103971115B (en) | Automatic extraction method for newly-increased construction land image spots based on NDVI and PanTex index | |
Zhao et al. | Automatic recognition of loess landforms using Random Forest method | |
Yang et al. | Automated extraction of building outlines from airborne laser scanning point clouds | |
CN109145992A (en) | Cooperation generates confrontation network and sky composes united hyperspectral image classification method | |
CN103440505B (en) | The Classification of hyperspectral remote sensing image method of space neighborhood information weighting | |
CN109766858A (en) | Three-dimensional convolution neural network hyperspectral image classification method combined with bilateral filtering | |
CN108345860A (en) | Personnel based on deep learning and learning distance metric recognition methods again | |
CN106023065A (en) | Tensor hyperspectral image spectrum-space dimensionality reduction method based on deep convolutional neural network | |
CN103208001A (en) | Remote sensing image processing method combined with shape self-adaption neighborhood and texture feature extraction | |
Ouma et al. | On the optimization and selection of wavelet texture for feature extraction from high‐resolution satellite imagery with application towards urban‐tree delineation | |
Sridharan et al. | Developing an object-based hyperspatial image classifier with a case study using WorldView-2 data | |
Zhang et al. | Mapping freshwater marsh species in the wetlands of Lake Okeechobee using very high-resolution aerial photography and lidar data | |
Aktas et al. | Landslide susceptibility mapping using an automatic sampling algorithm based on two level random sampling | |
CN103729652A (en) | Sparsity preserving manifold embedding based hyperspectral remote sensing image classification method | |
CN116030352B (en) | Long-time-sequence land utilization classification method integrating multi-scale segmentation and super-pixel segmentation | |
Li et al. | Incorporating open source data for Bayesian classification of urban land use from VHR stereo images | |
Li et al. | Integrating multiple textural features for remote sensing image change detection | |
CN109784209A (en) | Utilize the high and cold mountain area accumulated snow extracting method of high-resolution remote sensing image | |
Foroutan et al. | Automatic mapping and geomorphometry extraction technique for crevasses in geodetic mass-balance calculations at Haig Glacier, Canadian Rockies | |
Seydi et al. | Binary hyperspectral change detection based on 3D convolution deep learning | |
CN107341449A (en) | A kind of GMS Calculation of precipitation method based on cloud mass changing features | |
An et al. | Object-oriented urban dynamic monitoring—A case study of Haidian district of Beijing |
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 |