CN110298211B - River network extraction method based on deep learning and high-resolution remote sensing image - Google Patents
River network extraction method based on deep learning and high-resolution remote sensing image Download PDFInfo
- Publication number
- CN110298211B CN110298211B CN201810234115.5A CN201810234115A CN110298211B CN 110298211 B CN110298211 B CN 110298211B CN 201810234115 A CN201810234115 A CN 201810234115A CN 110298211 B CN110298211 B CN 110298211B
- Authority
- CN
- China
- Prior art keywords
- river
- sample
- remote sensing
- matrix
- sensing 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
Links
Images
Classifications
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/182—Network patterns, e.g. roads or rivers
Abstract
The invention discloses a river network extraction method based on deep learning and high-resolution remote sensing images, which comprises the following steps: acquiring a multi-scene high-resolution two-number multi-spectral remote sensing image, and dividing the remote sensing image into a training sample and a detection sample; remote sensing image preprocessing, wherein only images of a 4 th wave band and a 2 nd wave band are adopted for subsequent processing; selecting river sample points and non-river sample points as training sample points; establishing a deep learning model, inputting a training sample into the deep learning model for training to obtain a trained deep learning model; dividing the remote sensing image of the river network area to be extracted by a grid division method to obtain a detection sample; carrying out classification and identification through the trained model; and generating a river network. The technical scheme of the invention can ensure that the extracted river network is more real, accurate, fine and continuous and is not interfered by factors such as mountain shadow, buildings and the like.
Description
Technical Field
The invention belongs to the field of remote sensing image analysis, artificial intelligence and hydrology technology intersection, relates to a hydrological information extraction technology, and particularly relates to a river network extraction method based on deep learning and high-resolution remote sensing images.
Background
The river network of the drainage basin is an important geographic element, is an important component forming a landform 'skeleton', and has an important role in digital drainage basin expression. The traditional river network extraction method mainly adopts Digital Elevation Model (DEM) extraction. However, the existing digital elevation model is susceptible to depression and plain areas in the process of automatically extracting the virtual river network, and a discontinuous or parallel wrong river network is generated.
In recent years, with the rapid development of remote sensing technology, the application of high-resolution remote sensing images is becoming more and more popular. Many scholars have conducted research on extracting water and rivers from remote sensing images, but there are few studies on extracting river networks from remote sensing images. The algorithm for extracting the water body and river information by using the remote sensing image mainly comprises methods of water body indexes, supervised classification, unsupervised classification, gray level images, decision tree classification, mathematical morphology and the like. However, the river extracted by these algorithms has some problems, such as that the extracted river is easily disturbed by the shadow of the mountain, it is difficult to extract a fine river, and the extracted river has a partial interruption.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides the river network extraction method based on the deep learning and high-resolution remote sensing images, so that the extracted river network is more real, accurate, fine and continuous.
The technical scheme provided by the invention is as follows:
a river network extraction method based on deep learning and high-resolution remote sensing images comprises the following steps:
step 1: and acquiring a multi-scene high-resolution second remote sensing image, and determining whether each scene is used as a training sample or a detection sample.
The high-resolution remote sensing image adopted by the invention is a high-resolution second-number (GF-2) satellite remote sensing image. The high resolution binary (GF-2) satellite is the first civil optical remote sensing satellite with spatial resolution better than 1 meter independently developed in China, is provided with two high resolution 1 meter panchromatic and 4 meter multispectral cameras, and has the characteristics of sub-meter spatial resolution, high positioning precision, quick attitude maneuvering capability and the like. The multispectral spectral range with 4m resolution is 0.45-0.52 μm, 0.52-0.59 μm, 0.63-0.69 μm and 0.77-0.89 μm, which are sequentially marked as 1 st, 2 nd, 3 th and 4 th wave bands.
Step 2: and (3) preprocessing the remote sensing image, which comprises orthorectification processing and combination of the 4 th-2 th wave bands into a false color image, storing the false color image in a tif format under ENVI software, so that the gray value of each pixel is 255 at most, and further normalizing, wherein the normalization method is to divide the gray value of each pixel by 255. The subsequent processing only adopts the images of the 4 th wave band and the 2 nd wave band;
and step 3: training sample preparation
(1) Input of training samples
On the remote sensing image, points on rivers, including large rivers and small rivers, are selected, for the large rivers, not only are sample points in the center of the rivers selected, but also sample points which belong to the rivers and are located at the boundaries of the rivers are selected, and for the small rivers, sample points on the small rivers are mainly selected. And recording the horizontal and vertical coordinates of the river sample point.
On the remote sensing image, non-river points including mountains, shadows, clouds, buildings, farmlands, roads, railways and the like are selected, the non-river sample points are not selected near a water body as much as possible, and horizontal and vertical coordinates of the non-river sample points are recorded.
On a remote sensing image, taking the abscissa and the ordinate of a training sample point as the center, selecting a front m/2-1 row to a back m/2 row and a left m/2-1 column to a right m/2 column of a central point to form an m multiplied by m square neighborhood, wherein m belongs to a positive integer and is an even number; using the 4 th band and the 2 nd band of the image, the m × m square neighborhood is an m × m × 2 tensor (determined by the coordinates of the sample point), which is the input data of a training sample.
(2) Output of training samples
The output of each training sample is either a river or non-river, which is a binary problem. Let the output of the river sample be [ 10 ] and the output of the non-river sample be [ 01 ].
And 4, step 4: test sample preparation
(1) Inputting a detection sample, wherein the detection sample is obtained by dividing a remote sensing image of a river network area to be extracted;
acquiring at least one scene remote sensing image as a remote sensing image of a river network area to be extracted, and obtaining a detection sample by segmenting the remote sensing image of the river network area to be extracted through a grid segmentation method; the grid segmentation method adopts a plurality of segmentation modes (including a 16 segmentation mode in specific implementation), and divides an image into a plurality of sub-matrixes with the size of m rows and m columns to obtain a plurality of m multiplied by 2 tensors; taking each m × m × 2 tensor as a detection sample;
(2) detecting the output of the sample
For the deep learning procedure to be performed, the output of the test samples is prepared, where the output of each test sample is set to [ 01 ]. Because river network detection is a binary problem, the output result of the prediction model should be two values, and the preset output result is [ 01 ], namely, the output result represents two output values.
And 5: establishing a deep learning model
(1) Model building
And (3) applying a TensorFlow framework and establishing a deep learning model by using a python language under Spyde software. The deep learning model has the structure of 1 input layer, 1 convolution layer, 1 pooling layer, 1 full-connection layer and 1 output layer.
(2) Training of models
And putting the input and output data of the training sample into a deep learning model for training, and storing the trained model.
(3) Application of models
And reading the stored model, and putting the input and output data of each detection sample into the trained model for recognition to obtain a recognition result.
Step 6: generating a river network
(1) For each segmentation mode, the recognition result of each detection sample is reduced to an m × m (e.g., 16 × 16) matrix, and each element value in the matrix is consistent with the recognition result, and is either 0 or 1.
(2) For each segmentation mode, the recognition results are combined into a complete large matrix according to the corresponding positions during segmentation.
(3) The identification results obtained by different segmentation modes are overlapped, namely large matrix addition is carried out, and elements in the matrix are replaced by 1 as long as the elements are larger than 0.
(4) And (4) corresponding the superposed matrix to the image, and replacing the gray value of the image with 0 if the identification result is non-river, wherein the gray value of the image of the river is unchanged as the identification result.
(5) Because the identification result of each detection sample cannot be completely correct, for the sample with the wrong identification, scattered points are seen from the image, and a clearer and more accurate river network can be obtained by an algorithm of deleting the small-area image.
(6) And comparing the extracted river network with the google map of the corresponding position to detect whether the extraction effect of the river network is ideal.
Compared with the prior art, the invention has the beneficial effects that:
compared with the common remote sensing image, the high-resolution remote sensing image has higher resolution, and can identify the thinner rivers by combining a depth learning method.
Compared with single image element point identification in the prior art, the square neighborhood is taken as an identification unit, comprehensive investigation can be carried out in a larger range, and the image identification result cannot be obtained when the river image is partially blocked, such as a bridge on a river, the river image is partially not clear enough, and the like. Similarly, non-rivers can be better identified, reducing interference.
Compared with the prior art, the method has the advantages that the artificial intelligence method based on deep learning is applied, larger data volume can be processed, the method is more intelligent, the threshold value does not need to be manually set, the method is completely self-adaptive, rivers can be better identified, and the extracted river network is more real and accurate.
Compared with the prior art, the post-processing technology is added, including a grid segmentation method and small-area deletion, and a clearer, continuous and more precise river network is obtained.
Compared with the prior art, river network extraction is not affected by plains, freeze thawing, cities and mountain shadows.
Drawings
FIG. 1 is a block flow diagram of the method of the present invention.
FIG. 2 is a schematic diagram of a mesh segmentation method;
wherein, F1-F16 respectively represent 16 division modes included in the grid division method provided by the invention; a represents a remote sensing image matrix of a river network to be extracted; a. the1x1Is an element of the 1 st row and 1 st column of the matrix A, A1x16、A16x1、A16x16Respectively, row 1, column 16, row 16, column 1 and row 16, column 16 of matrix a.
Fig. 3 is a high-resolution second remote sensing image from which a river network needs to be extracted in the embodiment of the present invention.
Fig. 4 shows the result of river network extraction in the embodiment of the present invention.
Fig. 5 is a google map of a corresponding position extracted by a river network in the embodiment of the invention.
Detailed Description
The invention will be further described by way of examples, without in any way limiting the scope of the invention, with reference to the accompanying drawings.
The invention provides a river network extraction method based on deep learning and high-resolution remote sensing images, which can be used for realizing fine and continuous extraction of river network information and improving the accuracy of river network information extraction and the true degree of extraction results.
Fig. 1 shows a flow of a river network extraction method based on deep learning and high-resolution remote sensing images according to the present invention; the specific implementation steps are as follows:
step 1: acquiring a high-resolution second remote sensing image; it is determined whether each scene is to be a training sample or a detection sample.
And acquiring a multi-scene high-resolution two-number multispectral remote sensing image as a material for training samples, wherein the multi-scene remote sensing image needs to comprise rivers, mountains, cities, farmlands, roads and railways. In addition, at least one remote sensing image is obtained as a material for the river network to be extracted. The high-resolution two-number multispectral remote sensing image is used instead of the panchromatic image.
Step 2: and (3) preprocessing the remote sensing image, which comprises orthorectification processing and combination of the 4 th-2 th wave bands into a false color image, storing the false color image in a tif format under ENVI software, so that the gray value of each pixel is 255 at most, and further normalizing, wherein the normalization method is to divide the gray value of each pixel by 255. The subsequent processing only adopts the images of the 4 th wave band and the 2 nd wave band;
and step 3: training sample preparation
(1) Input of training samples
And selecting points located in rivers, including large rivers and small rivers, on the remote sensing image. And selecting non-river points including mountains, shadows, clouds, buildings, farmlands, roads, railways and the like, and recording the horizontal and vertical coordinates of the sample points.
On the remote sensing image, the horizontal and vertical coordinates of a sample point are taken as the center, the front 7 lines to the back 8 lines and the left 7 columns to the right 8 columns of the center point are selected to form a 16 × 16 square neighborhood, and because two wave bands are used, each sample is 16 × 16 × 2 tensor, which is input data of a training sample.
(2) Output of training samples
The output of each training sample is either a river or non-river, which is a binary problem. Let the output of the river sample be [ 10 ], the output of the non-river sample be [ 01 ],
and 4, step 4: test sample preparation
(1) Input of test sample
Selecting a remote sensing image of a river network area to be extracted, and dividing the image into 16 x 2 tensors by adopting a grid division method, wherein the 16 x 2 tensors are input data of a detection sample. The mesh segmentation method includes 16 segmentation modes.
(2) Detecting the output of the sample
For the deep learning procedure to be performed, the output of the test samples is prepared, where the output of each test sample is set to [ 01 ].
And 5: establishing a deep learning model
(1) Model building
And (3) applying a TensorFlow framework and establishing a deep learning model by using a python language under Spyde software. The deep learning model has the structure of 1 input layer, 1 convolution layer, 1 pooling layer, 1 full-connection layer and 1 output layer.
(2) Training of models
And putting the input and output data of the training sample into a deep learning model for training, and storing the trained model.
(3) Application of models
And reading the stored model, and putting the input and output data of each detection sample into the trained model for recognition to obtain a recognition result.
Step 6: generating a river network
(1) For each segmentation method, the recognition result of each detection sample is restored to a 16 × 16 matrix, and each element value in the matrix is consistent with the recognition result, and is either 0 or 1.
(2) For each segmentation method, the recognition results are combined into a complete large matrix according to the corresponding positions during segmentation.
(3) The identification results obtained by different segmentation methods are overlapped, namely large matrix addition is carried out, and elements in the matrix are replaced by 1 as long as the elements are larger than 0.
(4) And (4) corresponding the superposed matrix to the image, replacing the image gray value with 0 if the identification result is non-river, and keeping the image gray value of the river unchanged if the identification result is non-river.
(5) Because the identification result of each detection sample cannot be completely correct, for the sample with the wrong identification, scattered points are seen from the image, and a clearer and more accurate river network can be obtained by an algorithm of deleting the small-area image.
(6) And comparing the extracted river network with the google map of the corresponding position to detect whether the extraction effect of the river network is ideal.
Example (b):
in order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Step 1: obtaining high-resolution second remote sensing image
The high-resolution second remote sensing image is from the water conservancy department.
Step 2: remote sensing image preprocessing
(1) And (5) performing orthorectification on each scene remote sensing image under ENVI software.
(2) Under ENVI software, the 4 th-2 th wave band of the image after the orthometric correction is combined into a false color image which is stored as a tif format, so that the gray value of the pixel is maximum 255, and the normalization is further carried out, wherein the normalization method is to divide the gray value of each pixel by 255. The subsequent processing only adopts the images of the 4 th wave band and the 2 nd wave band;
and step 3: training sample preparation
(1) Opening the remote sensing image under the ENVI software, selecting points located in the river, including the center of a large river, the boundary of the large river and the center of a small river, and recording the horizontal and vertical coordinates of the points in Excel. In this example 20000 river sample points are selected.
(2) Opening the remote sensing image under the ENVI software, selecting non-river points including mountains, shadows, clouds, buildings, farmlands, roads, railways and the like, and recording the horizontal and vertical coordinates of the points in Excel. In this example 20000 non-river sample points are selected.
(3) Under Matlab software, two wave bands of No. 4 and No. 2 are selected, a 16 × 16 square neighborhood matrix is generated by taking the horizontal and vertical coordinates of a sample point as the center, and because two wave bands are used, each sample is a 16 × 16 × 2 tensor which is input data of a training sample.
(4) The output of the training samples is classified into two categories, so that the output of the river sample is recorded as [ 10 ], the output of the non-river sample is recorded as [ 01 ], and the output values of all the training samples are generated under Matlab software.
And 4, step 4: test sample preparation
(1) Determining a remote sensing image of a river network area to be extracted as a research object, in order to simplify calculation, a part of the image preprocessed in the step 2 is cut out, and an image with the size of 1600 × 1600 is selected here, as shown in fig. 3. The longitude and latitude coordinates of the junction of the two rivers are 114 degrees, 28 minutes, 52.91 seconds and 27 degrees, 7 minutes and 58.37 seconds.
(2) An image of 1600 × 1600 is divided into 10000 detection samples, each of which is a 16 × 16 × 2 tensor.
(3) For the needs of later deep learning program operation, there is also a need to prepare the output of the test sample, i.e. generate a 10000 row 2 column matrix, each row [ 01 ], just for the needs of program operation, and not to represent practical significance.
(4) In order to obtain a more complete river network, the invention adopts a plurality of segmentation modes for segmenting the image, which is called as a grid segmentation method, for example, a 16 segmentation mode is adopted in fig. 2. As shown in fig. 2, each division method is named by its starting point and is denoted as division methods F1 to F16. Points F1 to F16 are 16 elements obtained by performing 4 × 4 averaging on the first sub-matrix of the remote sensing image of the river network area to be extracted.
In the 1600 × 1600 matrix of the present embodiment, the division F1 indicates that 100 rows are equally divided from the point F1 downward, each row has a width of 16, and 100 columns are equally divided from the point F1 rightward, each column has a width of 16. In this way, the division method F1 obtains 100 × 100 — 10000 submatrices, each of which has a size of a square matrix of 16 rows and 16 columns, and each of which is a single detection sample, thereby being a tensor of 16 × 16 × 2 for each detection sample. The division pattern F11 is to discard all rows above the row where the point F11 is located (upper 7 rows in the submatrix), discard all columns to the left of the row where the point F11 is located (left 7 columns in the submatrix), the row and column where the point F11 is located are not discarded, divide the rows 99 equally from the point F11 downward, the width of each row is 16, divide the rows 99 equally from the point F11 rightward, and the width of each column is 16. Finally, the bottom most remaining row and the right most remaining column are discarded. In this way, the division method F11 provides sub-matrices of 99 × 99 to 9801, each having a size of a square matrix of 16 rows and 16 columns. Other division modes are analogized in the same way.
And 5: establishing a deep learning model
(1) Model building
And (3) applying a TensorFlow framework and establishing a deep learning model by using a python language under Spyde software. The deep learning model (the convolutional neural network is adopted for specific implementation) has the structure of 1 input layer, 1 convolutional layer, 1 pooling layer, 1 full-link layer and 1 output layer.
(2) Training of models
And putting input and output data of training samples into a deep learning model for training, wherein the training is performed for 30000 times totally, 50 samples are randomly selected for each training, and the model is stored after the training is finished.
(3) Application of models
And reading the stored model, putting input data of the detection samples obtained in a 16-division mode into the trained model for classification and identification, wherein the output result is two numerical values which respectively represent the probability that the detection sample belongs to a river or a non-river, and if the probability is high, the detection sample belongs to the river or the non-river, so that the classification result of each detection sample is obtained. For convenience of the following description of generating the river network, the result obtained by identifying the detection sample is recorded as 1 for the river and 0 for the non-river.
Step 6: generating a river network
(1) Each test sample yielded a 1 or 0 recognition result. If the number is 1, reducing the matrix into a 16 × 16 matrix, wherein each element of the matrix is 1; if 0, it is restored to a zero matrix of 16 × 16. And combining the recognition results into a complete large matrix according to the corresponding positions during the segmentation.
The 1 st division method F1 is combined into a 0, 1 matrix of 1600 × 1600, and the other 15 division methods are combined into a 0, 1 matrix of (16 × 99) × (16 × 99), and the 0, 1 matrix of 1600 × 1600 is obtained by filling the boundaries around with 0 according to the corresponding positions at the time of division.
Each segmentation method thus yields a corresponding 1600 x 1600 matrix of 0, 1. The 0, 1 matrixes obtained by the 16 division method are superposed and summed, and the elements in the matrixes are replaced by 1 as long as the elements are larger than 0, so that a comprehensive 1600 x 1600 0, 1 matrix is obtained. According to experience, the results of 16 times are not necessarily all superposed, and sometimes, only a plurality of 16 results may be better. This step can be implemented by Matlab software.
(2) The comprehensive 1600 x 1600 0 and 1 matrix obtained by identifying the detection sample corresponds to the 1600 x 1600 remote sensing image intercepted after preprocessing, and if the corresponding position is 0, the gray value of the image is replaced by 0; if the value is 1, the gray scale value of the image is kept unchanged.
Specifically, the obtained comprehensive 1600 × 1600 0 and 1 matrix has the same size as the matrix of the 1600 × 1600 remote sensing image intercepted after preprocessing, and can correspond to the matrix by the horizontal and vertical coordinates. Comparing elements in two matrixes at the same position, recording a matrix of the remote sensing image as A, recording a 0-1 matrix of the identification result as B, and if B (i, j) is 0, then A (i, j) is 0; if B (i, j) is not equal to 0, then A (i, j) remains unchanged.
Because the recognition result may not be complete and accurate, for the sample with the wrong recognition, scattered points are seen from the image, a better river network map can be obtained by deleting the small-area image, and finally the river network map is stored in the tif format. This step is implemented under Matlab software.
(3) And opening the extracted remote sensing image of the river network under the ENVI software, and viewing the extraction result of the river network, as shown in figure 4.
(4) According to the longitude and latitude of the detection sample, a map corresponding to the geographic position is found on the google map, as shown in fig. 5, the river network extraction method provided by the invention can be used for extracting a more continuous river network and extracting a finer river than that displayed on the google map in China at present.
It is noted that the disclosed embodiments are intended to aid in further understanding of the invention, but those skilled in the art will appreciate that: various substitutions and modifications are possible without departing from the spirit and scope of the invention and appended claims. Therefore, the invention should not be limited to the embodiments disclosed, but the scope of the invention is defined by the appended claims.
Claims (10)
1. A river network extraction method based on deep learning and high-resolution remote sensing images comprises the following steps:
step 1: acquiring a multi-scene high-resolution two-number multi-spectral remote sensing image, and dividing the remote sensing image into a training sample and a detection sample;
the spectral ranges of the 1 st, 2 nd, 3 th and 4 th wave bands of the high-resolution second-grade multispectral remote sensing image are respectively 0.45-0.52 mu m, 0.52-0.59 mu m, 0.63-0.69 mu m and 0.77-0.89 mu m;
step 2: remote sensing image preprocessing, including orthorectification processing and combining the 4 th-2 nd wave band into a false color image, and further normalizing, wherein the subsequent processing only adopts the images of the 4 th wave band and the 2 nd wave band;
and step 3: preparing a training sample;
31) acquiring training sample points as model input; the following operations are performed:
selecting river sample points and non-river sample points as training sample points on the training sample remote sensing image, and recording the abscissa and the ordinate of the training sample points;
on a remote sensing image, taking the abscissa and the ordinate of a training sample point as the center, selecting a front m/2-1 row to a back m/2 row and a left m/2-1 column to a right m/2 column of a central point to form an m multiplied by m square neighborhood, wherein m belongs to a positive integer and is an even number; adopting the 4 th wave band and the 2 nd wave band of the image, wherein the m multiplied by m square neighborhood is an m multiplied by 2 tensor as a training sample;
32) setting model output;
the recognition result of each training sample is river or non-river, and the output result is two numerical values respectively representing the probability that the sample belongs to the river and the probability that the sample belongs to the non-river; recording the output of the river sample as [ 10 ] and the output of the non-river sample as [ 01 ];
33) establishing a deep learning model, inputting training samples into the deep learning model for classification training to obtain a trained deep learning model;
and 4, step 4: preparing a detection sample;
41) obtaining at least one remote sensing image as a remote sensing image of a river network area to be extracted, and obtaining a detection sample by segmenting the remote sensing image of the river network area to be extracted through a grid segmentation method; the grid segmentation method adopts a plurality of segmentation modes to segment an image into a plurality of sub-matrixes with the size of m rows and m columns to obtain a plurality of m multiplied by 2 tensors; taking each m × m × 2 tensor as a detection sample;
42) carrying out classification and identification through the trained model;
putting each detection sample data into a trained model for classification and identification to obtain an identification result, and identifying to obtain a river or a non-river, wherein the river or the non-river is represented as 1 or 0;
and 5: generating a river network, and performing the following operations:
51) reducing the identification result of each detection sample into an m multiplied by m matrix, wherein each element value in the matrix is consistent with the identification result, the matrix element values belonging to the river classification are all 1, and the matrix element values belonging to the non-river classification are all 0;
52) aiming at each segmentation mode, combining the recognition results according to the corresponding positions in the segmentation process to obtain a combined complete large matrix;
53) adding the combined large matrixes obtained by different segmentation modes to obtain a superposition matrix; if the element value in the superposition matrix is greater than 0, replacing the value of the element with 1;
54) mapping the superposition matrix to the image; if the recognition result is non-river, replacing the image gray value with 0; if the recognition result is a river, the image gray value is unchanged; thereby obtaining a river network display image;
through the steps, the river network is extracted and displayed based on the deep learning and high-resolution remote sensing images.
2. The river network extraction method according to claim 1, wherein for the river network display image obtained in the step 54), scattered points in the image are deleted by an algorithm for deleting the small-area image, so as to remove the sample with the identification error.
3. The river network extraction method according to claim 1, wherein the step 2) stores the false color map in tif format under ENVI software, so that the gray value of the pixel is 255 at maximum, and further normalizes by dividing the gray value of each pixel by 255.
4. The river network extraction method according to claim 1, wherein step 31) is to obtain training sample points on the training sample remote sensing image, and the river sample points are obtained by specifically selecting points on a river; rivers include large rivers and small rivers; the large river sample points comprise a sample point positioned in the center of the river and a sample point positioned on the boundary of the river and belonging to the river; the non-river sample points include sample points of mountains, shadows, clouds, buildings, farmlands, roads, railways far away from the water body.
5. The river network extraction method according to claim 1, wherein step 31) selects the first 7 rows to the last 8 rows and the left 7 columns to the right 8 columns of the center point to form a 16 x 16 square neighborhood by taking the abscissa and the ordinate of the training sample point as the center, and each training sample is a 16 x 2 tensor; step 41) dividing the remote sensing image of the river network area to be extracted into a plurality of 16-row and 16-column sub-matrixes to obtain a plurality of 16 x 2 tensors; taking each 16 × 16 × 2 tensor as a detection sample; step 51) restoring the identification result of each detection sample to a matrix of 16 × 16.
6. The river network extraction method according to claim 1, wherein step 31) is implemented by establishing a deep learning model, specifically by using a TensorFlow framework and a Spyde software and using a python language; the structure of the deep learning model comprises an input layer, a convolution layer, a pooling layer, a full-link layer and an output layer.
7. The river network extraction method according to claim 1, wherein the step 41) obtains the detection sample by a mesh segmentation method, and the mesh segmentation method comprises a 16-segmentation method.
8. The river network extraction method according to claim 7, wherein the 16 division modes are F1-F16; the segmentation mode F1 specifically includes: assuming that the remote sensing image data is a square matrix of n rows and n columns, wherein n is km, n, k and m all belong to positive integers, equally dividing k rows from a point F1 to the lower part, wherein the width of each row is m, equally dividing k columns from a point F1 to the right part, and the width of each column is m; obtaining k × k sub-matrixes, wherein the size of each sub-matrix is a square matrix of m rows and m columns, and each sub-matrix is a detection sample;
points F1 to F16 are 16 elements obtained by performing 4 × 4 averaging on the first sub-matrix of the remote sensing image of the river network area to be extracted.
9. The river network extraction method according to claim 8, wherein the division pattern F11 is to discard all rows above the row of point F11 and to discard all columns to the left of the row of point F11; equally dividing k-1 rows from the row where the point F11 is located and the column to the bottom, wherein the width of each row is m, equally dividing k-1 columns from the point F11 to the right, and the width of each column is m; finally, abandoning the lowest remaining row and the rightmost remaining column; thereby obtaining (k-1) × (k-1) sub-matrices, each having a size of m rows and m columns of square matrix.
10. The river network extraction method according to claim 7, wherein the segmentation mode F1 is specifically as follows: starting from a point F1, equally dividing 100 rows downwards, wherein the width of each row is 16, starting from a point F1, equally dividing 100 columns to the right, and the width of each column is 16; 10000 sub-matrixes are obtained, the size of each sub-matrix is a square matrix with 16 rows and 16 columns, and each sub-matrix is a detection sample; the division mode F11 is to discard all rows above the row of point F11 and to discard all columns to the left of the row of point F11; equally dividing 99 rows from the row where the point F11 is located and 99 columns from the right, wherein the width of each row is 16, and equally dividing 99 columns from the point F11, and the width of each column is 16; finally, abandoning the lowest remaining row and the rightmost remaining column; this yields 99 × 99-9801 sub-matrices, each having a size of 16 rows and 16 columns of square matrices.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810234115.5A CN110298211B (en) | 2018-03-21 | 2018-03-21 | River network extraction method based on deep learning and high-resolution remote sensing image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810234115.5A CN110298211B (en) | 2018-03-21 | 2018-03-21 | River network extraction method based on deep learning and high-resolution remote sensing image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110298211A CN110298211A (en) | 2019-10-01 |
| CN110298211B true CN110298211B (en) | 2021-03-23 |
Family
ID=68025336
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810234115.5A Active CN110298211B (en) | 2018-03-21 | 2018-03-21 | River network extraction method based on deep learning and high-resolution remote sensing image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110298211B (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110992330B (en) * | 2019-11-28 | 2021-05-25 | 桂林理工大学 | Multi-level integral relaxation matching high-resolution ortho-image shadow detection under artificial shadow drive |
| CN111178230B (en) * | 2019-12-26 | 2022-03-15 | 武汉大学 | Intelligent extraction method for river beach in remote sensing image |
| CN111738168A (en) * | 2020-06-24 | 2020-10-02 | 中水北方勘测设计研究有限责任公司 | Satellite image river two-side sand production extraction method and system based on deep learning |
| CN112084843B (en) * | 2020-07-28 | 2024-03-12 | 北京工业大学 | Multispectral river channel remote sensing monitoring method based on semi-supervised learning |
| CN112785521A (en) * | 2021-01-19 | 2021-05-11 | 澜途集思生态科技集团有限公司 | Remote sensing image processing method under haze condition |
| CN113033386B (en) * | 2021-03-23 | 2022-12-16 | 广东电网有限责任公司广州供电局 | High-resolution remote sensing image-based transmission line channel hidden danger identification method and system |
| CN113888670B (en) * | 2021-10-18 | 2022-05-20 | 珠江水利委员会珠江水利科学研究院 | Method for generating high-precision two-dimensional terrain of one-way river channel based on deep learning |
| CN114155253B (en) * | 2021-12-09 | 2023-04-07 | 电子科技大学 | Connection method for strip-shaped element fracture based on remote sensing image segmentation |
| CN116630811B (en) * | 2023-06-07 | 2024-01-02 | 自然资源部国土卫星遥感应用中心 | River extraction method, river extraction device, terminal equipment and readable storage medium |
| CN117392539B (en) * | 2023-10-13 | 2024-04-09 | 哈尔滨师范大学 | River water body identification method based on deep learning, electronic equipment and storage medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102855490A (en) * | 2012-07-23 | 2013-01-02 | 黑龙江工程学院 | Object-neural-network-oriented high-resolution remote-sensing image classifying method |
| CN104462494A (en) * | 2014-12-22 | 2015-03-25 | 武汉大学 | Remote sensing image retrieval method and system based on non-supervision characteristic learning |
| CN106127784A (en) * | 2016-07-01 | 2016-11-16 | 辽宁工程技术大学 | A kind of high-resolution remote sensing image dividing method |
| 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 |
| CN107610141A (en) * | 2017-09-05 | 2018-01-19 | 华南理工大学 | A kind of remote sensing images semantic segmentation method based on deep learning |
| CN107688818A (en) * | 2016-08-05 | 2018-02-13 | 中国电力科学研究院 | A kind of path intelligent selecting method and system based on satellite remote-sensing image signature analysis |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105303198B (en) * | 2015-11-17 | 2018-08-17 | 福州大学 | A kind of remote sensing image semisupervised classification method learnt from fixed step size |
| CN105931179B (en) * | 2016-04-08 | 2018-10-26 | 武汉大学 | A kind of image super-resolution method and system of joint sparse expression and deep learning |
| CN107247938B (en) * | 2017-06-08 | 2019-12-06 | 中国科学院遥感与数字地球研究所 | high-resolution remote sensing image urban building function classification method |
| CN107729922A (en) * | 2017-09-20 | 2018-02-23 | 千寻位置网络有限公司 | Remote sensing images method for extracting roads based on deep learning super-resolution technique |
-
2018
- 2018-03-21 CN CN201810234115.5A patent/CN110298211B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102855490A (en) * | 2012-07-23 | 2013-01-02 | 黑龙江工程学院 | Object-neural-network-oriented high-resolution remote-sensing image classifying method |
| CN104462494A (en) * | 2014-12-22 | 2015-03-25 | 武汉大学 | Remote sensing image retrieval method and system based on non-supervision characteristic learning |
| CN106127784A (en) * | 2016-07-01 | 2016-11-16 | 辽宁工程技术大学 | A kind of high-resolution remote sensing image dividing method |
| CN107688818A (en) * | 2016-08-05 | 2018-02-13 | 中国电力科学研究院 | A kind of path intelligent selecting method and system based on satellite remote-sensing image signature analysis |
| 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 |
| CN107610141A (en) * | 2017-09-05 | 2018-01-19 | 华南理工大学 | A kind of remote sensing images semantic segmentation method based on deep learning |
Non-Patent Citations (2)
| Title |
|---|
| 高光谱遥感影像分类研究进展;杜培军 等;《遥感学报》;20160325;全文 * |
| 高分辨率遥感影像目标分类与识别研究进展;刘扬 等;《地球信息科学学报》;20150828;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110298211A (en) | 2019-10-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110298211B (en) | River network extraction method based on deep learning and high-resolution remote sensing image | |
| CN111986099B (en) | Tillage monitoring method and system based on convolutional neural network with residual error correction fused | |
| Khandelwal et al. | An approach for global monitoring of surface water extent variations in reservoirs using MODIS data | |
| CN109409263B (en) | Method for detecting urban ground feature change of remote sensing image based on Siamese convolutional network | |
| Cheng et al. | A spatial and temporal nonlocal filter-based data fusion method | |
| US9942440B2 (en) | Image-based field boundary detection and identification | |
| Li et al. | A 30-year (1984–2013) record of annual urban dynamics of Beijing City derived from Landsat data | |
| CN111738165B (en) | Method for extracting individual plant canopy from high-resolution unmanned aerial vehicle visible light remote sensing image | |
| CN111626947B (en) | Map vectorization sample enhancement method and system based on generation of countermeasure network | |
| CN111640159A (en) | Remote sensing image change detection method based on twin convolutional neural network | |
| Abd et al. | Maximum likelihood for land-use/land-cover mapping and change detection using Landsat satellite images: A case study “South of Johor” | |
| CN111028255A (en) | Farmland area pre-screening method and device based on prior information and deep learning | |
| CN111046613B (en) | Optimal river channel calculation method based on path tracking and river network extraction method based on multi-temporal remote sensing image | |
| Ermolaev et al. | Automated construction of the boundaries of basin geosystems for the Volga Federal District | |
| CN113887515A (en) | Remote sensing landslide identification method and system based on convolutional neural network | |
| CN109389062A (en) | Utilize the method for High Resolution Spaceborne SAR image zooming-out lake land and water cut-off rule | |
| CN111881816A (en) | Long-time-sequence river and lake ridge culture area monitoring method | |
| Grigillo et al. | Automated building extraction from IKONOS images in suburban areas | |
| CN115294460B (en) | Method for determining degradation degree of phyllostachys praecox forest, medium and electronic device | |
| Wahbi et al. | A deep learning classification approach using high spatial satellite images for detection of built-up areas in rural zones: Case study of Souss-Massa region-Morocco | |
| Xia et al. | Reconstructing Missing Information of Remote Sensing Data Contaminated by Large and Thick Clouds Based on an Improved Multitemporal Dictionary Learning Method. | |
| CN114494821A (en) | Remote sensing image cloud detection method based on feature multi-scale perception and self-adaptive aggregation | |
| CN115512247A (en) | Regional building damage grade assessment method based on image multi-parameter extraction | |
| CN116012723A (en) | Wetland type extraction method and device based on time sequence remote sensing image and electronic equipment | |
| Wan et al. | Mapping annual urban change using time series Landsat and NLCD |
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 |