CN111798461B - Pixel-level remote sensing image cloud area detection method for guiding deep learning by coarse-grained label - Google Patents

Pixel-level remote sensing image cloud area detection method for guiding deep learning by coarse-grained label Download PDF

Info

Publication number
CN111798461B
CN111798461B CN202010563344.9A CN202010563344A CN111798461B CN 111798461 B CN111798461 B CN 111798461B CN 202010563344 A CN202010563344 A CN 202010563344A CN 111798461 B CN111798461 B CN 111798461B
Authority
CN
China
Prior art keywords
cloud
remote sensing
convolution
activation
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.)
Active
Application number
CN202010563344.9A
Other languages
Chinese (zh)
Other versions
CN111798461A (en
Inventor
李彦胜
陈蔚
张永军
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University WHU
Original Assignee
Wuhan University WHU
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Wuhan University WHU filed Critical Wuhan University WHU
Priority to CN202010563344.9A priority Critical patent/CN111798461B/en
Publication of CN111798461A publication Critical patent/CN111798461A/en
Application granted granted Critical
Publication of CN111798461B publication Critical patent/CN111798461B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • 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

Abstract

The invention discloses a pixel-level remote sensing image cloud area detection method based on deep learning, which is based on the realistic problem that a cloud area detection method based on deep learning needs large cost of marking, and comprises the steps of firstly training a depth network model with good robustness and generating a cloud activation map by using easily-obtained image block-level labels and combining a local pooling layer pruning strategy and a global convolution pooling layer under the constraint of a remote sensing image block data set of the coarse-grained image block-level labels, and then obtaining a final cloud mask map through threshold segmentation. The invention can greatly reduce the marking work, simultaneously realize the pixel-level accurate cloud area detection of the remote sensing image, and effectively improve the efficiency and the performance of the cloud area detection of the remote sensing image.

Description

Pixel-level remote sensing image cloud area detection method for guiding deep learning by coarse-grained label
Technical Field
The invention belongs to the technical field of remote sensing and artificial intelligence, relates to a remote sensing image cloud area detection method for deep learning, and particularly relates to a pixel-level remote sensing image cloud area detection method for deep learning guided by coarse-grained labels.
Background
Cloud region detection is a key problem in remote sensing image interpretation and application, and a large amount of cloud cover influences usability of remote sensing image data and increases difficulty in remote sensing image interpretation. Cloud region detection detects a cloud-containing region in a remote sensing image through various methods in the fields of remote sensing and computer vision, and saves transmission bandwidth and storage space and reduces waste of resources by not issuing images containing more clouds on the aspect of on-satellite application; and on the aspect of ground application, data preparation is provided for applications such as cloud removal and image recovery, subsequent large-range continuous mapping, dynamic monitoring based on remote sensing images and the like.
In recent years, a large number of cloud region detection methods based on artificial structural features have been proposed in academia. With the continuous development of the related art, the method based on deep learning is also used for solving the cloud area detection problem in a large quantity. In general, the performance of a remote sensing image cloud area detection task can be remarkably improved by the deep learning-based method. However, the superior performance of deep learning relies on a large number of accurate pixel-level labels, and the labeling of labels is time-consuming and labor-intensive. Considering that different types of satellites often have great difference in spectral and spatial resolutions, for each remote sensing satellite image, a deep learning-based method needs a corresponding pixel-level labeling data set to operate, and further needs a great deal of manpower for labeling. Therefore, it is of great significance to explore an advanced cloud area detection method capable of reducing the labeling workload.
Disclosure of Invention
The invention provides a pixel-level remote sensing image cloud area detection method based on deep learning, which aims to solve the problem that a large amount of marking cost is needed in a cloud area detection method based on deep learning.
The technical scheme adopted by the invention is as follows: a pixel-level remote sensing image cloud region detection method for deep learning guided by coarse-grained labels comprises the following steps:
step 1: inputting a remote sensing image data set D { (b) with a coarse-grained labeln，yn) 1,2, …, N, where b isnRepresenting the nth remote sensing image block, y, of the data set DnThe coarse-grained remote sensing image block level label corresponding to the nth remote sensing image block is represented, and N represents the total number of the remote sensing image blocks in the data set D;
step 2: learning hyper-parameters of a deep convolutional network model on the data set D in the step 1, wherein the hyper-parameters comprise a convolution weight C, a global convolution pooling weight G and a cloud activation weight W;
and step 3: inputting an image I needing cloud area detection, and calculating a cloud activation map MI of the image I by using the I, the convolution weight C, the global convolution pooling weight G and the cloud activation weight W in the step 2;
and 4, step 4: cloud of image I in step 3Activation map MIPerforming threshold segmentation to calculate the cloud mask image S of the image II；
Further, y in step 1nHas two forms, yn＝[1，0]Representing the nth remote sensing image block b in the data set DnThe label of (A) is cloud-containing, yn＝[0，1]Representing the nth remote sensing image block b in the data set DnThe tag of (a) is cloud-free;
further, the specific implementation of step 2 includes the following sub-steps:
step 2.1: one remote sensing image block b of the data set D in the step 1nInputting the feature map into a deep convolutional network model, and outputting a feature map according to the following formula:
Figure GDA0003511105510000021
wherein f isnAn output feature map representing the last convolutional layer in the deep convolutional network model.
Figure GDA0003511105510000022
Representing the integral representation of operations such as convolution, activation operation and the like in the deep convolutional network model, and C represents the convolution weight in the deep convolutional network model;
step 2.2: for step 2.1 fnThe global convolution pooling is carried out on each channel, and the activation value of each channel is calculated, wherein the formula is as follows:
Figure GDA0003511105510000023
wherein the content of the first and second substances,
Figure GDA0003511105510000024
denotes fnThe (c) th channel of (a),
Figure GDA0003511105510000025
represents
Figure GDA0003511105510000026
Activation value, G, at the k channel after global convolution poolingKE G represents the weight of the global convolution pooling weight G at the kth channel,
Figure GDA0003511105510000027
represents fnPerforming spatial convolution operation with each channel of G;
step 2.3: the hyper-parameters C, G, W of the deep convolutional network model are learned using a cross-entropy loss function based on softmax. The formula is as follows:
Figure GDA0003511105510000028
wherein the content of the first and second substances,
Figure GDA0003511105510000031
representing that the cloud activation weight value is in the kth channel, and the weight value of the category c is; class c corresponds to bnLabel of (a), yn＝[0，1]When c is 1, yn＝[1，0]When c is 0; d represents the total number of channels.
Step 2.4: and (3) repeating the step 2.1 to the step 2.3 for each remote sensing image block in the data set D until all data participate in deep convolutional network model training, repeating iteration for 10 times until the network converges, and acquiring a deep convolutional network model and a hyper-parameter thereof: a convolution weight C, a global convolution pooling weight G and a cloud activation weight W;
further, the specific implementation of step 3 includes the following sub-steps:
step 3.1: inputting an image I needing cloud region detection, and cutting the image I into overlapped image blocks { a ] by using a sliding window algorithm1，a2，…，am}；
Step 3.2: for a certain image block a in step 3.1, inputting the certain image block a into the deep convolutional network model in step 2.4 to output a characteristic map f, wherein the formula is as follows:
Figure GDA0003511105510000032
wherein f represents the output characteristic diagram of the last convolution layer in the deep convolution network model.
Figure GDA0003511105510000033
Representing the overall representation of operations such as convolution, activation operations and the like in the deep convolutional network model;
step 3.3: for the kth band f of the output feature f in step 3.2kCalculating the adjusted feature map TkThe formula is as follows:
Figure GDA0003511105510000034
wherein, TkIs the k-th band fkAnd (5) adjusting the characteristic diagram.
Figure GDA0003511105510000035
Is calculated using f in equation 2.1kActivation value at k channel, GKE G represents the weight of the global convolution pooling weight G obtained in the step 2.4 in the k channel, tau (f)k) Is fkA statistical value of (a), including a mean or median;
step 3.4: repeating the steps 3.2 and 3.3 for each channel of the image block a in the step 3.2, and calculating a cloud activation map MaThe formula is as follows:
Figure GDA0003511105510000036
wherein the content of the first and second substances,
Figure GDA0003511105510000037
k is 1,2, …, d represents the value of the cloud activation weight in the k channel;
step 3.5: for all image blocks { a ] in step 3.11，a2，…，amRepeating the steps 3.2, 3.3 and 3.4, and calculating the cloud activation corresponding to each image blockDrawing (A)
Figure GDA0003511105510000038
The cloud activation image M of the image I can be obtained after all the cloud activation images are splicedI。
Further, the specific implementation of step 4 includes the following sub-steps:
step 4.1: for all image blocks b in the data set D in step 1 without clouds1 -，b2 -，…，bt -And calculating a cloud activation image corresponding to each image block
Figure GDA0003511105510000041
Calculate out
Figure GDA0003511105510000042
The mean μ and standard deviation σ of;
step 4.2: calculating a segmentation threshold h by using the mean value mu and the standard deviation sigma of the cloud activation map in the step 4.1, wherein the formula is as follows:
h ═ μ + kxoσ (formula seven);
wherein k is a coefficient;
step 4.3: cloud activation map M of video I using segmentation threshold h calculated in step 4.2IPerforming threshold segmentation to calculate the cloud mask image S of the image IIThe formula is as follows:
Figure GDA0003511105510000043
wherein, (i, j) is a cloud activation map MIOr cloud mask map SIThe horizontal and vertical coordinates of (1);
furthermore, the network structure of the deep convolutional network model contains 10 convolutional layers, 1 global convolutional pooling layer, 1 full-link layer and 1 softmax classification layer.
Furthermore, the size of a convolution kernel in the convolution layer is 3 × 3, the sliding step of convolution is 1 × 1, and a ReLU nonlinear active layer is connected behind the convolution layer.
Further, the window size of the global convolution pooling layer is 230 × 230.
Further, in step 4.2, k is 0.7.
The invention has the following advantages: the local pooling layer pruning strategy used in the method can greatly improve the resolution of the network output characteristic diagram, and can be applied to other tasks with high requirements on the resolution of the characteristic diagram, such as small target detection and the like; the global pooling convolutional layer used in the method can better extract the spatial variation information in the feature map, thereby improving the quality of the output feature map. Compared with the existing cloud detection method, the method can complete the training of the depth network only by the image block-level label, and greatly reduces the required labeling cost while realizing accurate pixel-level cloud detection.
Drawings
FIG. 1 is a schematic diagram of a deep convolutional network structure according to an embodiment of the present invention;
FIG. 2 is a diagram illustrating a global convolutional pooling operation according to an embodiment of the present invention; wherein (a) is a forward propagation process and (b) is a backward propagation process.
FIG. 3 is a schematic diagram of a cloud activation map generation process according to an embodiment of the present invention;
FIG. 4 is a schematic diagram illustrating a process of generating a cloud activation map for a test image according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of a cloud detection result generation process according to an embodiment of the present invention; the method comprises the following steps of (a) obtaining an original image, (b) obtaining a reference cloud area image, (c) obtaining a cloud activation image, and (d) obtaining a cloud mask image.
Detailed Description
In order to facilitate the understanding and implementation of the present invention for those of ordinary skill in the art, the present invention is further described in detail with reference to the accompanying drawings and examples, it is to be understood that the embodiments described herein are merely illustrative and explanatory of the present invention and are not restrictive thereof.
The invention provides a remote sensing image scene classification method based on fault-tolerant deep learning, which comprises the following steps of:
step 1: inputting remote sensing image data set with coarse-grained labelD＝{(bn，yn) 1,2, …, N, where b isnRepresenting the nth remote sensing image block in the data set D; y isnCoarse-grained remote-sensing image block-level label, y, corresponding to nth remote-sensing image blocknHas two forms, yn＝[1，0]Representing the nth remote sensing image block b in the data set DnThe label of (A) is cloud-containing, yn＝[0，1]Representing the nth remote sensing image block b in the data set DnThe tag of (a) is cloud-free; n represents the total number of remotely sensed image blocks in data set D.
Step 2: and (3) learning hyper-parameters of the deep convolutional network model on the data set D in the step 1, wherein the hyper-parameters comprise a convolution weight C, a global convolution pooling weight G and a cloud activation weight W. The method specifically comprises the following substeps:
please see fig. 1, step 2.1: one remote sensing image block b of the data set D in the step 1nInputting the feature map into a deep convolutional network model, and outputting a feature map according to the following formula:
Figure GDA0003511105510000051
wherein f isnAn output feature map representing the last convolutional layer in the deep convolutional network model.
Figure GDA0003511105510000052
Represents the overall representation of operations such as convolution, activation operations and the like in the deep convolutional network model, and C represents the convolution weight in the deep convolutional network model.
Please see fig. 2, step 2.2: for step 2.1 fnThe global convolution pooling is carried out on each channel, and the activation value of each channel is calculated, wherein the formula is as follows:
Figure GDA0003511105510000053
wherein the content of the first and second substances,
Figure GDA0003511105510000054
denotes fnThe (c) th channel of (a),
Figure GDA0003511105510000055
represents
Figure GDA0003511105510000056
Activation value, G, at the k channel after global convolution poolingKE G represents the weight of the global convolution pooling weight G at the kth channel,
Figure GDA0003511105510000057
represents fnAnd the spatial convolution operation between the G channels.
Step 2.3: the hyper-parameters C, G, W of the deep convolutional network model are learned using a cross-entropy loss function based on softmax. The formula is as follows:
Figure GDA0003511105510000061
wherein the content of the first and second substances,
Figure GDA0003511105510000062
representing that the cloud activation weight value is in the kth channel, and the weight value of the category c is; class c corresponds to bnLabel of (a), yn＝[0，1]When c is 1, yn＝[1，0]When c is 0; d represents the total number of channels.
Step 2.4: and (3) repeating the step 2.1 to the step 2.3 for each remote sensing image block in the data set D until all data participate in deep convolutional network model training, repeating iteration for 10 times until the network converges, and acquiring a deep convolutional network model and a hyper-parameter thereof: a convolution weight C, a global convolution pooling weight G, and a cloud activation weight W.
And step 3: inputting an image I needing cloud area detection, and calculating a cloud activation map M of the image I by using the I, the convolution weight C, the global convolution pooling weight G and the cloud activation weight W in the step 2I. The method specifically comprises the following substeps:
step 3.1: inputting an image I needing cloud region detection, and segmenting the image I into overlapped images by using a sliding window algorithmImage block { a1，a2，…，am}。
Please see fig. 3, step 3.2: for a certain image block a in step 3.1, inputting the certain image block a into the deep convolutional network model in step 2.4 to output a characteristic map f, wherein the formula is as follows:
Figure GDA0003511105510000063
wherein f represents the output characteristic diagram of the last convolution layer in the deep convolution network model.
Figure GDA0003511105510000064
Representing an overall representation of the operations of convolution, activation operations, etc. in a deep convolutional network model.
Step 3.3: for the kth band f of the output feature f in step 3.2kCalculating the adjusted feature map TkThe formula is as follows:
Figure GDA0003511105510000065
wherein, TkIs the k-th band fkAnd (5) adjusting the characteristic diagram.
Figure GDA0003511105510000066
Is calculated using f in equation 2.1kActivation value at k channel, GKE G represents the weight of the global convolution pooling weight G obtained in the step 2.4 in the k channel, tau (f)k) Is fkA statistical value such as a mean or median value.
Step 3.4: repeating the steps 3.2 and 3.3 for each channel of the image block a in the step 3.2, and calculating a cloud activation map MaThe formula is as follows:
Figure GDA0003511105510000067
wherein，
Figure GDA0003511105510000068
k is 1,2, …, d represents the value of the cloud activation weight value on the k channel, d represents the total channel number
Please see fig. 4, step 3.5: for all image blocks { a ] in step 3.11，a2，…，amRepeating the steps 3.2, 3.3 and 3.4, and calculating a cloud activation graph corresponding to each image block
Figure GDA0003511105510000071
The cloud activation image M of the image I can be obtained after all the cloud activation images are splicedI。
Please see fig. 5, step 4: cloud activation map M for image I in step 3IPerforming threshold segmentation to calculate the cloud mask image S of the image II. The method specifically comprises the following substeps:
step 4.1: for all image blocks b in the data set D in step 1 without clouds1 -，b2 -，…，bt -Repeating the steps 3.2, 3.3 and 3.4, and calculating a cloud activation graph corresponding to each image block
Figure GDA0003511105510000072
Calculate out
Figure GDA0003511105510000073
Mean μ and standard deviation σ of.
Step 4.2: calculating a segmentation threshold h by using the mean value mu and the standard deviation sigma of the cloud activation map in the step 4.1, wherein the formula is as follows:
h ═ μ + kxoσ (formula seven);
wherein k is an empirical coefficient obtained through experiments.
Step 4.3: using the segmentation threshold h calculated in step 4.2 to the cloud activation map M of the image I obtained in step 3.5IPerforming threshold segmentation to calculate the cloud mask image S of the image IIThe formula is as follows:
Figure GDA0003511105510000074
wherein, (i, j) is a cloud activation map MIOr cloud mask map SIThe abscissa and ordinate of (a).
TABLE 1 network architecture configuration of deep convolutional network model used in the method
Figure GDA0003511105510000075
Figure GDA0003511105510000081
Table 1 shows the network structure of the deep convolutional network model used in the method, and the size of the input image processed by the network structure is 500 × 500 × 4. In table 1, "convolution kernel" specifies the size of the convolution kernel reception field, the dimension dim of the input data, and the number num of convolution kernels, which are expressed as size × size × dim × num by a formula; "step size" means the sliding step size of the convolution; "ReLU nonlinear activation" means that a ReLU nonlinear activation layer is connected after the convolution layer; the "window size" represents the window size of the global convolution pooling layer. As shown in table 1, the network structure contains 10 convolutional layers, 1 global convolutional pooling layer, 1 fully-connected layer and 1 softmax classification layer.
In order to analyze the effect of the experimental coefficient k in step 4.2 on the deep learning, table 2 shows the performance indexes of the method under the settings of different coefficients k. At k 0.7, the process achieves the best performance.
TABLE 2 several performance indexes of the method under different coefficients k
Figure GDA0003511105510000082
It should be understood that parts of the specification not set forth in detail are well within the prior art.
It should be understood that the above description of the preferred embodiments is given for clarity and not for any purpose of limitation, and that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (7)

1. A pixel-level remote sensing image cloud region detection method for deep learning guided by coarse-grained labels is characterized by comprising the following steps:
step 1: inputting a remote sensing image data set D { (b) with a coarse-grained labeln,yn) 1,2, …, N, where b isnRepresenting the nth remote sensing image block, y, of the data set DnThe coarse-grained remote sensing image block level label corresponding to the nth remote sensing image block is represented, and N represents the total number of the remote sensing image blocks in the data set D;
step 2: learning hyper-parameters of a deep convolutional network model on the data set D in the step 1, wherein the hyper-parameters comprise a convolution weight C, a global convolution pooling weight G and a cloud activation weight W;
the specific implementation of the step 2 comprises the following substeps:
step 2.1: one remote sensing image block b of the data set D in the step 1nInputting the feature map into a deep convolutional network model, and outputting a feature map according to the following formula:
Figure FDA0003511105500000011
wherein f isnAn output characteristic diagram representing the last convolutional layer in the deep convolutional network model;
Figure FDA0003511105500000012
representing the integral representation of convolution and activation operation in the deep convolution network model, and C representing the convolution weight in the deep convolution network model;
step 2.2: to the step of2.1 in fnThe global convolution pooling is carried out on each channel, and the activation value of each channel is calculated, wherein the formula is as follows:
Figure FDA0003511105500000013
wherein the content of the first and second substances,
Figure FDA0003511105500000014
denotes fnThe (c) th channel of (a),
Figure FDA0003511105500000015
represents
Figure FDA0003511105500000016
Activation value, G, at the k channel after global convolution poolingKE G represents the weight of the global convolution pooling weight G at the kth channel,
Figure FDA0003511105500000017
represents fnPerforming spatial convolution operation with each channel of G;
step 2.3: learning the hyper-parameters C, G, W of the deep convolutional network model using a cross entropy loss function based on softmax, the formula is as follows:
Figure FDA0003511105500000018
wherein the content of the first and second substances,
Figure FDA0003511105500000019
representing that the cloud activation weight value is in the kth channel, and the weight value of the category c is; class c corresponds to bnLabel of (a), yn＝[0,1]When c is 1, yn＝[1,0]When c is 0; d represents the total number of channels;
step 2.4: and (3) repeating the step 2.1 to the step 2.3 for each remote sensing image block in the data set D until all data participate in deep convolutional network model training, repeating iteration for a plurality of times until the network converges, and acquiring a deep convolutional network model and a hyper-parameter thereof: a convolution weight C, a global convolution pooling weight G and a cloud activation weight W;
and step 3: inputting an image I needing cloud area detection, and calculating a cloud activation map M of the image I by using the I, the convolution weight C, the global convolution pooling weight G and the cloud activation weight W in the step 2I；
The specific implementation of the step 3 comprises the following substeps:
step 3.1: inputting an image I needing cloud region detection, and cutting the image I into overlapped image blocks { a ] by using a sliding window algorithm1,a2,…,am}；
Step 3.2: and (3) inputting a certain image block a in the step 3.1 into the deep convolution network model to output a characteristic map f, wherein the formula is as follows:
Figure FDA0003511105500000021
wherein f represents the output characteristic diagram of the last convolution layer in the deep convolution network model,
Figure FDA0003511105500000022
representing an overall representation of convolution, activation arithmetic operations in a deep convolutional network model;
step 3.3: for the kth band f of the output feature f in step 3.2kCalculating the adjusted feature map TkThe formula is as follows:
Figure FDA0003511105500000023
wherein, TkIs the k-th band fkAfter the feature map is adjusted,
Figure FDA0003511105500000024
is to use step 2Equation 1 for equation fkActivation value at k channel, GKE G represents the weight of the global convolution pooling weight G obtained in the step 2.4 in the k channel, tau (f)k) Is fkA statistical value of (a), including a mean or median;
step 3.4: repeating the steps 3.2 and 3.3 for each channel of the image block a in the step 3.2, and calculating a cloud activation map MaThe formula is as follows:
Figure FDA0003511105500000025
wherein the content of the first and second substances,
Figure FDA0003511105500000026
representing the value of the cloud activation weight in the kth channel;
step 3.5: for all image blocks { a ] in step 3.11,a2,…,amRepeating the steps 3.2, 3.3 and 3.4, and calculating a cloud activation graph corresponding to each image block
Figure FDA0003511105500000027
Splicing all the cloud activation maps to obtain a cloud activation map M of the image II；
And 4, step 4: cloud activation map M for image I in step 3IPerforming threshold segmentation to calculate the cloud mask image S of the image II。
2. The pixel-level remote sensing image cloud region detection method for coarse-grained label guided deep learning according to claim 1, characterized in that: y in step 1nHas two forms, yn＝[1,0]Representing the nth remote sensing image block b in the data set DnThe label of (A) is cloud-containing, yn＝[0,1]Representing the nth remote sensing image block b in the data set DnThe tag of (1) is cloud-free.
3. The pixel-level remote sensing image cloud region detection method for coarse-grained label guided deep learning according to claim 1, characterized in that: the specific implementation of the step 4 comprises the following substeps:
step 4.1: all the cloud-free image blocks in the data set D in step 1 are recorded as { b }1 -,b2 -,…,bt -And calculating a cloud activation image corresponding to each image block
Figure FDA0003511105500000031
Calculate out
Figure FDA0003511105500000032
The mean μ and standard deviation σ of;
step 4.2: calculating a segmentation threshold h by using the mean value mu and the standard deviation sigma of the cloud activation map in the step 4.1, wherein the formula is as follows:
h ═ μ + kxoσ (formula seven);
wherein k is a coefficient;
step 4.3: cloud activation map M of video I using segmentation threshold h calculated in step 4.2IPerforming threshold segmentation to calculate the cloud mask image S of the image IIThe formula is as follows:
Figure FDA0003511105500000033
wherein, (i, j) is a cloud activation map MIOr cloud mask map SIThe abscissa and ordinate of (a).
4. The pixel-level remote sensing image cloud region detection method for coarse-grained label guided deep learning according to claim 1, characterized in that: the network structure of the deep convolutional network model comprises 10 convolutional layers, 1 global convolutional pooling layer, 1 full-link layer and 1 softmax classification layer.
5. The pixel-level remote sensing image cloud region detection method for coarse-grained label guided deep learning according to claim 4, characterized in that: the size of a convolution kernel in the convolution layer is 3 multiplied by 3, the sliding step length of convolution is 1 multiplied by 1, and a ReLU nonlinear activation layer is connected behind the convolution layer.
6. The pixel-level remote sensing image cloud region detection method for coarse-grained label guided deep learning according to claim 4, characterized in that: the window size of the global convolution pooling layer is 230 × 230.
7. The pixel-level remote sensing image cloud region detection method for coarse-grained label guided deep learning according to claim 3, characterized in that: in step 4.2, k is 0.7.
CN202010563344.9A 2020-06-19 2020-06-19 Pixel-level remote sensing image cloud area detection method for guiding deep learning by coarse-grained label Active CN111798461B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010563344.9A CN111798461B (en) 2020-06-19 2020-06-19 Pixel-level remote sensing image cloud area detection method for guiding deep learning by coarse-grained label

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010563344.9A CN111798461B (en) 2020-06-19 2020-06-19 Pixel-level remote sensing image cloud area detection method for guiding deep learning by coarse-grained label

Publications (2)

Publication Number Publication Date
CN111798461A CN111798461A (en) 2020-10-20
CN111798461B true CN111798461B (en) 2022-04-01

Family

ID=72804040

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010563344.9A Active CN111798461B (en) 2020-06-19 2020-06-19 Pixel-level remote sensing image cloud area detection method for guiding deep learning by coarse-grained label

Country Status (1)

Country Link
CN (1) CN111798461B (en)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109063663A (en) * 2018-08-10 2018-12-21 武汉大学 A kind of spissatus detection of timing remote sensing image and minimizing technology by slightly to essence
CN110119728A (en) * 2019-05-23 2019-08-13 哈尔滨工业大学 Remote sensing images cloud detection method of optic based on Multiscale Fusion semantic segmentation network

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9721181B2 (en) * 2015-12-07 2017-08-01 The Climate Corporation Cloud detection on remote sensing imagery
CN108805861A (en) * 2018-04-28 2018-11-13 中国人民解放军国防科技大学 Remote sensing image cloud detection method based on deep learning
CN108932474B (en) * 2018-05-28 2022-03-15 北京航空航天大学 Remote sensing image cloud judgment method based on full convolution neural network composite characteristics
CN109255294A (en) * 2018-08-02 2019-01-22 中国地质大学（北京） A kind of remote sensing image clouds recognition methods based on deep learning
CN109740639B (en) * 2018-12-15 2021-02-19 中国科学院深圳先进技术研究院 Wind cloud satellite remote sensing image cloud detection method and system and electronic equipment
CN109801293B (en) * 2019-01-08 2023-07-14 平安科技（深圳）有限公司 Remote sensing image segmentation method and device, storage medium and server
CN109934200B (en) * 2019-03-22 2023-06-23 南京信息工程大学 RGB color remote sensing image cloud detection method and system based on improved M-Net
CN111274865B (en) * 2019-12-14 2023-09-19 深圳先进技术研究院 Remote sensing image cloud detection method and device based on full convolution neural network

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109063663A (en) * 2018-08-10 2018-12-21 武汉大学 A kind of spissatus detection of timing remote sensing image and minimizing technology by slightly to essence
CN110119728A (en) * 2019-05-23 2019-08-13 哈尔滨工业大学 Remote sensing images cloud detection method of optic based on Multiscale Fusion semantic segmentation network

Also Published As

Publication number Publication date
CN111798461A (en) 2020-10-20

Similar Documents

Publication Publication Date Title
CN108573276B (en) Change detection method based on high-resolution remote sensing image
CN110287849B (en) Lightweight depth network image target detection method suitable for raspberry pi
CN112308860B (en) Earth observation image semantic segmentation method based on self-supervision learning
CN111340738B (en) Image rain removing method based on multi-scale progressive fusion
CN112132149B (en) Semantic segmentation method and device for remote sensing image
CN111461258A (en) Remote sensing image scene classification method of coupling convolution neural network and graph convolution network
CN112132058A (en) Head posture estimation method based on multi-level image feature refining learning, implementation system and storage medium thereof
CN112257727B (en) Feature image extraction method based on deep learning self-adaptive deformable convolution
CN115984850A (en) Lightweight remote sensing image semantic segmentation method based on improved Deeplabv3+
CN113269224A (en) Scene image classification method, system and storage medium
CN111723660A (en) Detection method for long ground target detection network
CN114943876A (en) Cloud and cloud shadow detection method and device for multi-level semantic fusion and storage medium
CN112001293A (en) Remote sensing image ground object classification method combining multi-scale information and coding and decoding network
CN104899821A (en) Method for erasing visible watermark of document image
CN113657387A (en) Semi-supervised three-dimensional point cloud semantic segmentation method based on neural network
CN115457311A (en) Hyperspectral remote sensing image band selection method based on self-expression transfer learning
CN114863266A (en) Land use classification method based on deep space-time mode interactive network
CN114596477A (en) Foggy day train fault detection method based on field self-adaption and attention mechanism
Wang et al. Multi‐scale network for remote sensing segmentation
Tian et al. Semantic segmentation of remote sensing image based on GAN and FCN network model
CN114092467A (en) Scratch detection method and system based on lightweight convolutional neural network
CN111723934B (en) Image processing method and system, electronic device and storage medium
CN111798461B (en) Pixel-level remote sensing image cloud area detection method for guiding deep learning by coarse-grained label
CN113052121A (en) Multi-level network map intelligent generation method based on remote sensing image
CN114708434A (en) Cross-domain remote sensing image semantic segmentation method based on adaptation and self-training in iterative domain

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