CN111028255B - Farmland area pre-screening method and device based on priori information and deep learning - Google Patents
Farmland area pre-screening method and device based on priori information and deep learning Download PDFInfo
- Publication number
- CN111028255B CN111028255B CN201811188060.5A CN201811188060A CN111028255B CN 111028255 B CN111028255 B CN 111028255B CN 201811188060 A CN201811188060 A CN 201811188060A CN 111028255 B CN111028255 B CN 111028255B
- Authority
- CN
- China
- Prior art keywords
- farmland
- areas
- area
- grid
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/13—Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- G06T7/60—Analysis of geometric attributes
- G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Geometry (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a farmland area pre-screening method and device based on priori information and deep learning, wherein the method comprises the following steps: dividing the satellite image according to grid division to form an image block area; removing non-farmland areas in the satellite images in advance based on multi-source priori information, and reserving farmland areas to form a preliminary screening result; extracting a farmland/non-farmland true value based on the high-precision farmland boundary; deep learning is carried out based on farmland/non-farmland truth values, image blocking areas in the primary screening result are classified, whether the current image blocking areas belong to farmland areas or not is judged, non-farmland areas are removed, farmland areas are reserved, and a pre-screening result is formed; and storing the pre-screening result of the farmland area in a grid mode. Through practical tests, the pre-screening method provided by the invention can reduce the total processing time for extracting the boundary of the farmland with high precision to below 30% of the original processing time.
Description
Technical Field
The invention relates to the technical field of farmland area screening, in particular to a farmland area pre-screening method and device based on priori information and deep learning.
Background
With the rapid development of satellite remote sensing technology, sensing technology and high-precision positioning technology, the coverage area of satellite remote sensing images is continuously enlarged, and the resolution and positioning precision are also continuously improved. Meanwhile, the satellite/minisatellite constellation is continuously formed, so that the revisiting time of satellite remote sensing to the same area is shorter and shorter, and some satellites even have revisiting capability of 1 to 2 days. The satellite images can realize large-area dynamic monitoring under medium-low resolution, can realize small-range accurate detection under high resolution, and are widely applied to various fields of domestic and foreign military and civil. Along with the massive acquisition of the massive high-resolution satellite remote sensing data, intelligent precise agricultural application becomes a hot spot and trend, such as unmanned aerial vehicle plant protection, intelligent pest and disease monitoring, crop growth monitoring and the like, and how to extract space distribution elements of farmlands and crops from the massive high-resolution satellite remote sensing data becomes a key technology for restricting precise agricultural development.
At present, the boundaries of farmland blocks (an independent farmland area with obvious edges in satellite images) required in the unmanned aerial vehicle plant protection process are mainly extracted manually on satellite/aerial images or obtained in a mapping/measuring mode on an operation site, so that the unmanned aerial vehicle plant protection system is low in efficiency and high in cost, and the centralized management and updating cannot be realized due to scattered data distribution. Therefore, the automatic and accurate extraction of the boundaries of farmland blocks is a key technology for accurate agricultural applications. In recent years, computer vision and machine learning have been rapidly developed, in which image semantic segmentation (semantic segmentation refers to marking all pixels in an image to obtain a plurality of homogeneous regions, so that all pixels in each region belong to the same type of ground object or target), and single object extraction-oriented instance segmentation (instance segmentation, which is a further extension of image semantic segmentation, can distinguish different object individuals belonging to the same type (such as adjacent different farmland blocks, adjacent two buildings, etc.), while distinguishing different object individuals belonging to the same type (such as farmland and architecture) according to the image semantic segmentation, and is called object instance), but these methods are complex in calculation, time-consuming and serious, and require about 30 seconds for processing a 0.3m resolution 8×2048 image according to a test, and require about 6 months for performing a national boundary extraction even in the case of computing power capable of processing 100 images in parallel. In practice, the cultivated land area of China only accounts for about 14% of the total area of the domestic soil, and high-precision farmland boundary extraction is not needed to be carried out on all areas. Therefore, in practical application, a rapid farmland pre-screening method is needed, which can reject non-farmland areas nationwide and reserve farmland areas, and based on the method, high-precision farmland boundary extraction can be performed on farmland area images, so that the processing time can be obviously saved.
The technical proposal close to or relevant to the invention is as follows:
1. accurate extraction method for unmanned aerial vehicle aerial image farmland block objects, publication number: CN107563413a. The invention provides a precise extraction method of unmanned aerial vehicle aerial image farmland block objects, which comprises the steps of firstly obtaining image edge information by utilizing contour detection based on spectrum information, then obtaining a bottom layer segmentation block by adopting watershed transformation, generating a multi-scale segmentation map based on contour intensity, and finally realizing non-farmland region rejection through supervised image classification.
Note that: the above method realizes high-precision farmland block boundary extraction, but the algorithm is complex, 30 seconds are required for processing single-frame images with the resolution of 2048×2048 pixels, and 6 months are required for finishing national farmland boundary extraction (under the condition of the computing power capable of processing 100 images in parallel), so that a rapid farmland region pre-screening method is required to perform high-precision boundary extraction only for images containing farmland regions.
2. Method and device for extracting water information, publication number: CN103793907B. The method utilizes digital elevation model DEM (Digital Elevation Model, digital high-level model) data of the appointed drainage basin to assist in extracting water information in remote sensing image data. First, DEM data is used to extract river information, identify linear rivers, and generate a river buffer. Then, first water body information of the appointed watershed is extracted based on the remote sensing image data. And finally, combining the river buffer area and the first water body information to generate final target water body information.
3. Extraction method and system of target water body, publication number: CN104463166a. The scheme provides a method for providing support for extracting the water body in the remote sensing image data by utilizing GIS vector data of the target water body. The method uses vector data of the geometric shape of the water body to generate a statistical histogram of the remote sensing image, and assists in judging the boundary between the water body and the land. Based on the statistical histogram and a preset threshold, the method judges boundary information, so that water body extraction is performed.
Note that: the two methods use DEM/GIS information to assist water extraction, the prior information is single, only the DEM or GIS information of the water is used, and the advantage that the prior information can be directly inquired is not used in the processing method, so that simplicity and effectiveness are achieved.
4. Hyperspectral remote sensing image ground object classification method based on overrun learning machine, publication number: CN106897737a. The method is a hyperspectral image ground object classification method based on an overrun learning machine (Extreme Learning Machine, ELM), a multichannel fusion network model (Enhanced Hierarchical Extreme Learning Machine, EH-ELM) is designed, and the characteristic information of a ground object target is obtained by utilizing the spatial characteristics and the spectral characteristics of the ground object target in an image, so that the hyperspectral image ground object classification problem is solved.
5. Deep learning-based high-resolution PolSAR image urban feature classification method, publication number: CN107194349a. According to the method, a deep learning neural network is constructed, the deep learning neural network is trained by collecting high-resolution PolSAR source data which are similar enough, characteristics which are more beneficial to classification of urban data are obtained through migration learning, and the PolSAR urban image data to be classified are introduced into the trained deep learning neural network to obtain classification results.
Note that: the two methods use a machine learning/deep learning method to classify the features of the hyperspectral image/PolSAR image, the data acquisition cost of the hyperspectral image/PolSAR image is high, the updating period is slow, and meanwhile, the adopted machine learning method needs a large number of training sample supports, and the true value acquisition cost is high.
The development of high-resolution satellite remote sensing technology and high-precision positioning technology enables accurate agricultural application to be a trend, and automatic accurate extraction of farmland boundaries is the basis of accurate agricultural application such as unmanned aerial vehicle plant protection, pest and disease monitoring and the like. In order to obtain a high-precision boundary, the existing high-precision farmland boundary extraction method mainly adopts complex semantic segmentation or instance segmentation, so that the method has complex processing process and long execution time.
Disclosure of Invention
The invention provides a farmland area pre-screening method and device based on priori information and deep learning aiming at high-resolution satellite images, which are used for removing non-farmland areas in a nationwide range and reserving images containing farmland areas for high-precision farmland boundary extraction, so that farmland boundary extraction efficiency is improved.
The technical scheme adopted by the invention is as follows:
a farmland area pre-screening method based on priori information and deep learning comprises the following steps:
dividing the satellite image according to grid division to form an image block area;
removing non-farmland areas in the satellite images in advance based on multi-source priori information, and reserving farmland areas to form a preliminary screening result;
extracting a farmland/non-farmland true value based on the high-precision farmland boundary;
deep learning is carried out based on farmland/non-farmland truth values, image blocking areas in the primary screening result are classified, whether the current image blocking areas belong to farmland areas or not is judged, non-farmland areas are removed, farmland areas are reserved, and a pre-screening result is formed;
and storing the pre-screening result of the farmland area in a grid mode.
Further, the removing of the non-farmland area in advance based on the multi-source priori information specifically comprises the following steps:
performing grid division on the candidate region to obtain grid units;
adopting a theme prior data set as prior information, wherein the theme prior data set comprises a grid data set and a vector data set;
removing grid cells of the non-region of interest based on the land boundary, and reserving the grid cells of the region of interest;
filtering the grid cells based on the grid data set to remove mountain and desert areas;
filtering the grid unit based on the vector data set to remove the water area;
grid cells after removing mountain, desert and water areas are used as candidate farmland areas.
Further, the specific steps of filtering the grid cells based on the raster data set are as follows:
matching the grid data set with the grid cells by adopting a mapping method to obtain target areas corresponding to the grid cells in the grid data set;
and counting the proportion of the pixels of which the types are mountain or desert in the target area to the pixels of the target area, and judging that the grid unit belongs to the mountain area or the desert area if the proportion is larger than a threshold value, or else, judging that the grid unit belongs to the candidate farmland area.
Further, the specific steps of filtering the grid cells based on the vector data set are as follows:
calculating the area of the grid cells;
calculating the ratio of the sum of the area of the grid unit and the area of the intersection of all the water polygons in the vector data set to the area of the grid unit;
and if the ratio is greater than the threshold value, judging that the grid unit belongs to the water area, otherwise, judging that the grid unit belongs to the candidate farmland area.
Further, based on the high-precision farmland boundary extraction, a farmland/non-farmland true value is obtained, and a farmland/non-farmland binary mask corresponding to the satellite image is generated, and the method specifically comprises the following steps of:
selecting satellite images, and extracting farmland boundaries of the satellite images with high precision;
and generating a farmland/non-farmland binary mask by using the farmland boundary, wherein an image area surrounded by the farmland boundary is a farmland, and an image area outside the farmland boundary is a non-farmland, so that a training truth value is provided for rapid classification of the farmland/non-farmland based on deep learning.
Further, the area ratio of the farmland area in each grid unit is estimated by adopting a regression model based on a convolutional neural network, so that the image blocking area is classified rapidly, and the method specifically comprises deep learning-based regression model training and deep learning-based regression model farmland/non-farmland rapid classification.
Further, the regression model training based on deep learning specifically includes the following steps:
the satellite image is segmented to form satellite image segmentation blocks;
based on the extracted farmland/non-farmland truth value, calculating a truth value cutting block corresponding to each satellite image cutting block in a farmland/non-farmland truth value diagram;
performing binarization processing on the true value cutting blocks, and calculating the pixel ratio of pixel points of the farmland area in each satellite image cutting block to serve as a training label of the satellite image cutting block;
and inputting the satellite image segmentation blocks and the corresponding training labels into a convolutional neural network for model training.
Further, the regression model farmland/non-farmland rapid classification based on deep learning specifically comprises the following steps:
cutting the satellite image to form cutting blocks;
inputting each segmentation block into a farmland/non-farmland rapid classification model based on a convolutional neural network, and giving out a farmland duty ratio estimated value of each segmentation block;
and (3) averaging the estimated values of all the cut blocks in the current satellite image, comparing the average value with a threshold value, if the average value is larger than the threshold value, considering that the current satellite image contains farmland areas, and if the average value is smaller than the threshold value, considering that the current satellite image does not contain farmland areas.
The invention also provides a farmland boundary extraction method, which comprises the farmland region pre-screening method based on priori information and deep learning.
The invention also provides a farmland area pre-screening device based on priori information and deep learning, which comprises:
the image segmentation unit is used for segmenting the satellite image according to grid division to form an image block area;
the primary screening unit is used for removing non-farmland areas in the satellite images in advance based on the multi-source priori information, reserving the farmland areas and forming a primary screening result;
the true value extraction unit is used for extracting farmland/non-farmland true values based on the high-precision farmland boundaries;
the pre-screening unit is used for performing deep learning based on farmland/non-farmland truth values, classifying image blocking areas in the primary screening result, judging whether the current image blocking areas belong to farmland areas, removing the non-farmland areas, and reserving the farmland areas to form a pre-screening result;
and the storage unit is used for storing the pre-screening result of the farmland area in a grid mode.
The present invention also provides a memory storing a computer program, the computer program being executable by a processor to:
dividing the satellite image according to grid division to form an image block area;
removing non-farmland areas in the satellite images in advance based on multi-source priori information, and reserving farmland areas to form a preliminary screening result;
extracting a farmland/non-farmland true value based on the high-precision farmland boundary;
deep learning is carried out based on farmland/non-farmland truth values, image blocking areas in the primary screening result are classified, whether the current image blocking areas belong to farmland areas or not is judged, non-farmland areas are removed, farmland areas are reserved, and a pre-screening result is formed;
and storing the pre-screening result of the farmland area in a grid mode.
Through practical tests, the non-farmland area accounting for more than 70% of the national area can be removed by using the pre-screening method provided by the invention, and the total processing time for extracting the farmland boundary with high precision is reduced to be less than 30% of the original processing time due to the fact that the time for pre-screening the farmland area for one image is very small (about 2 seconds on average), namely, the time for extracting the farmland boundary nationally is reduced to be less than 2 months (under the condition of the computing capacity capable of processing 100 images in parallel).
Drawings
FIG. 1 is a general flow chart of farmland area pre-screening based on prior information and deep learning;
FIG. 2 is a flow chart of a large-scale non-farmland fast culling based on multi-source prior information;
FIG. 3 is a regression model training flow diagram based on deep learning;
FIG. 4 is a flow chart of a regression model farmland/non-farmland fast classification based on deep learning;
fig. 5 is a block diagram of a farmland area pre-screening device based on prior information and deep learning.
Detailed Description
Firstly, adopting a multisource GIS (Geographic Information System ) and DEM theme priori data (such as land boundary, mountain, desert and water area) to pre-reject a large-scale non-farmland area, and reserving the farmland area; then further removing more non-farmland areas through deep learning, wherein farmland/non-farmland truth values required by the deep learning are extracted through high-precision farmland boundaries; finally, the pre-screening information of the farmland area is stored in a grid mode, and in practical application, the high-precision farmland boundary extraction efficiency is improved by processing the image only containing the farmland area. The invention is further described below with reference to the drawings and examples.
Embodiment one:
the farmland area pre-screening overall flow based on priori information and deep learning is shown in fig. 1, and comprises the following steps:
step 1, large-scale non-farmland rapid elimination based on multi-source priori information: the method has the advantages that the rapid elimination of large-area water bodies, mountain lands and desert gobi areas is realized by utilizing the theme GIS information of the water bodies (including sea-land boundaries), mountain lands, desert gobi and the like;
step 2, farm/non-farm truth value generation based on high-precision farm boundary extraction: the farmland boundary information extracted based on the high-precision farmland boundary can generate a farmland/non-farmland binary mask corresponding to an image, and a low-cost farmland/non-farmland training true value is provided for rapid classification of farmland/non-farmland based on deep learning;
step 3, large-scale farmland/non-farmland rapid classification based on deep learning: through deep learning, the rapid ground object classification of the image blocking areas is realized, and the integral discrimination of the current image belonging to farmlands or non-farmlands is realized through the category statistical information of a plurality of image blocking areas in one image;
step 4, storing farmland area pre-screening results in a grid mode: based on multisource prior information, mainly rejecting mountain areas, water bodies and other non-farmland areas, based on deep learning, forest areas, buildings and other areas can be further rejected, and a final national farmland area pre-screening result is obtained by combining two pre-screening information in a nationwide grid division mode.
The method and the device finally obtain whether the area corresponding to each grid cell in the whole country is a farmland or non-farmland area, wherein white color in the pre-screening result diagram shown in fig. 1 indicates that the current grid cell is a farmland area, and black color indicates that the current grid cell is a non-farmland area. And the farmland area obtained based on pre-screening is subjected to high-precision farmland boundary extraction, so that the time consumption in the treatment process can be obviously reduced.
The farmland area pre-screening method based on priori information and deep learning is further described below:
1. multi-source priori information-based large-range non-farmland rapid rejection
The invention realizes the rapid elimination of large-area water bodies, mountain lands, desert gobi areas by utilizing the subject GIS information of water bodies (including sea-land dividing lines), mountain lands, desert gobi and the like. The large-scale non-farmland quick reject flow based on multi-source prior information is shown in fig. 2. Based on grid division of the whole country, the corresponding non-farmland part in the grid is removed by utilizing a large-scale theme GIS data set (such as mountain land, water area and the like).
The core task of non-farmland culling using a priori data sets is to mark each grid cell in the future as a non-farmland area based on a priori information. The invention uses 0 to represent that the grid unit is a non-farmland; 1 represents that the grid unit is a farmland, and the concrete steps are as follows:
step 11, a large-scale theme priori data set is adopted as priori information, wherein the theme priori data set comprises theme data sets of mountain, desert and water areas: the mountain dataset records most mountain and mountain geographic information; the desert data set records geographic information of the desert and partial barren lands; the water area data set record is geographical information of water surface distribution of rivers, lakes and the like. In addition, the invention also uses the Chinese land boundary to filter out grids outside the Chinese land range.
And step 12, removing part of grid cells except the grid China by adopting the China land boundary information, and only reserving the grid cells in China.
And 13, filtering the grid cells by adopting a mountain and desert grid subject grid data set, and removing mountain and desert parts. The invention adopts a mapping method to match the subject raster data with the grid cells, thereby eliminating the grid cells corresponding to the mountain and the desert area. The mapping method is used for finding a target (mountain or desert) area corresponding to the grid unit in the grid data, and comprises the following specific steps:
step 131, obtaining geographic coordinates of two points of the upper left corner and the lower right corner of the grid unit, wherein the geographic coordinates are represented by (lon 1, lat 1) and (lon 2, lat 2);
step 132, calculating the pixel coordinates corresponding to the two geographic coordinates in the raster data by combining the geographic reference information of the raster data, wherein the pixel coordinates are represented by (x 1, y 1) and (x 2, y 2);
step 133, taking two points (x 1, y 1) and (x 2, y 2) as the top left corner and bottom right corner vertices to obtain a rectangular area, namely a target area corresponding to the grid unit in the grid data;
and step 134, counting the proportion of the total pixels of the target area occupied by the pixels with the types of mountain (or desert) in the target area, if the proportion is more than 85%, judging that the grid unit belongs to the mountain (or desert) area, otherwise, belonging to the candidate farmland area.
And 14, further processing the grid cells by adopting the water area vector data set, and screening out cells in the water area range in the grid. The invention adopts a method for calculating the intersection area of two polygons to calculate the cells falling in the water area, and the specific steps are as follows:
step 141, for a grid cell, calculating its area S;
step 142, calculating the intersecting area Sk of the grid unit and the kth water polygon in the water area vector data set, wherein k=1-K, and K is the total number of the water polygons in the water area vector data set;
step 143, calculating the ratio of the sum of the intersecting areas of the grid unit and all the polygons of the water body to the area of the grid unit, namely
And 144, if the ratio is greater than 85%, judging that the grid unit belongs to the water area, otherwise, judging that the grid unit belongs to the candidate farmland area.
Finally, grid cells for removing mountain, desert and water areas are obtained as candidate farmland areas.
2. Farm/non-farm truth generation based on high-precision farm boundary extraction
Based on farmland boundary information extracted from high-precision farmland boundaries, a farmland/non-farmland binary mask corresponding to an image can be generated, and a low-cost farmland/non-farmland training true value is provided for rapid classification of farmland/non-farmland based on deep learning. The method comprises the following specific steps:
step 21, selecting a certain number of satellite images, and extracting the farmland boundaries with high precision;
and 22, generating a farmland/non-farmland binary mask by using the farmland boundary, wherein an image area surrounded by the farmland boundary is farmland, and an image area outside the farmland boundary is non-farmland, and providing a training truth value for rapid classification of farmland/non-farmland based on deep learning.
The high-precision farmland boundary extraction module is used for extracting the farmland/non-farmland boundaries of the farmland, so that the high-precision farmland boundary extraction module can extract the farmland boundaries accurately, and the time consumption of farmland boundary extraction is saved.
3. Deep learning-based large-range farmland/non-farmland rapid classification
The invention adopts a regression model based on deep learning to realize rapid ground object classification of the image blocking areas, and realizes the integral discrimination of the current image belonging to farmlands or non-farmlands through the category statistical information of a plurality of image blocking areas in one image. Because the high-resolution satellite images are organized in a grid mode, the actual coverage area of one image corresponding to each grid reaches 0.36 square kilometer, and if whether farmland is contained in each grid area is evaluated in a conventional classification mode, the actual occupation ratio of the farmland area in the grid cannot be embodied. The invention adopts a regression model based on a deep convolutional neural network to evaluate the area ratio of a farmland area in each grid, and specifically comprises two parts of regression model training based on deep learning and regression model farmland/non-farmland rapid classification based on deep learning:
1) Regression model training based on deep learning
The training process of the farmland/non-farmland rapid classification model based on the deep convolutional neural network is shown in fig. 3, and the specific steps are as follows:
step 311, dividing a high-resolution satellite image according to a specified rule to obtain divided blocks x 1 ，x 2 ，...x n Wherein x is i The i-th cut block is represented, and the total number of cut blocks is n. The splitting method includes, but is not limited to, non-overlapping equal-size splitting according to a grid (for example, splitting blocks are formed by using a grid method every 256×256 pixels), overlapping equal-size splitting according to a grid, random equal-size splitting, non-equal-size splitting and the like.
Step 312, obtaining a true farmland/non-farmland training value of the current high-resolution satellite image by utilizing high-precision farmland boundary extraction; calculating the dividing block x of each high-resolution satellite image i Corresponding cut-segment z in farmland/non-farmland truth-chart i Wherein i represents a dicing sequence number;
step 313, dividing the truth into blocks z i Binarization processing is carried out, farmland area is assigned to 1, non-farmland area is assigned to 0, and truth value cutting block z is calculated i The number of pixels (i.e. farmland area) with a median value of 1 is the pixel number in each cut block to be the ratio y i This is used as the high-resolution satellite image segmentation block x i Is used for training the label;
step 314, training data set d= { (x) 1 ，y 1 )，(x 2 ，y 2 )，...，(x n ，y n ) The high-score satellite image is segmented and the corresponding labels are input into a convolutional neural network, and the characteristic expression f (x) i )＝wx i +b, where w and b are network parameters such that the network output f (x i ) As close as possible to tag y i And solving the network parameters w and b by adopting a method of minimizing the mean square error, thereby realizing model training.
2) Regression model farmland/non-farmland rapid classification based on deep learning
The flow of rapid classification of farmland/non-farmland by adopting a regression model based on a convolutional neural network is shown in fig. 4, and the specific steps are as follows:
step 321, dividing a high-resolution satellite image into a plurality of blocks q 1 ，q 2 ，...q m Wherein q is j Representing the j-th block, m representing the total number of blocks, one implementation is to segment the current image into a number of blocks of equal size (e.g., 256×256 pixels) that do not overlap each other;
step 322, divide each block q j Inputting a farmland/non-farmland rapid classification model based on a convolutional neural network, and giving out a farmland duty ratio estimated value p of each block j ；
Step 323, divide all blocks q in the current image j Is estimated to be p j And solving the average value P, comparing the average value P with a threshold value T, and considering that a certain farmland area is contained in the current image if the average value P is larger than the threshold value, and considering that the farmland area is not contained in the current image if the average value P is smaller than the threshold value, wherein the threshold value T=0.25 is adopted in the invention.
4. Storing farmland area pre-screening results in a grid mode
Based on the grid mode, the farmland pre-screening result (namely farmland/non-farmland information corresponding to a certain geographic area) is stored, so that linear search is conveniently performed by utilizing the longitude and latitude positions, the subsequent high-precision farmland boundary extraction module is used as priori information, and meanwhile continuous updating of farmland pre-screening information is facilitated. The flow of storing the farmland area pre-screening result in a grid mode is as follows: 1) Firstly, grid division is needed to be carried out on the whole country, the invention adopts grid units with the size of 600m multiplied by 600m to carry out the grid division on the whole country, and meanwhile, the satellite images are cut according to the grid division, namely, one grid corresponds to a satellite image with the resolution of 0.3 with the size of 2048 multiplied by 2048 pixels; 2) Based on the defined grids, most mountainous regions, water bodies and the like can be removed by utilizing large-scale non-farmland rapid removal based on multi-source prior information; 3) Based on the farmland area grids obtained by screening and satellite images corresponding to the grids, most forest areas, building areas and the like can be further removed by utilizing large-range farmland/non-farmland rapid classification based on deep learning, and a final national farmland area pre-screening result is obtained.
Because the high-precision farmland boundary extraction module is time-consuming, farmland area information obtained by farmland pre-screening is utilized to guide the high-precision farmland boundary extraction module, so that the high-precision farmland boundary extraction module only processes images containing farmland areas, and the farmland boundary extraction time consumption can be obviously reduced (through actual tests, the total farmland boundary extraction processing time is reduced to be less than 30% of the original processing time).
The invention also provides a farmland boundary extraction method, which comprises a farmland area pre-screening method based on priori information and deep learning.
Embodiment two:
the invention also provides a farmland area pre-screening device based on priori information and deep learning, as shown in fig. 5, comprising:
the image segmentation unit is used for segmenting the satellite image according to grid division to form an image block area;
the primary screening unit is used for removing non-farmland areas in the satellite images in advance based on the multi-source priori information, reserving the farmland areas and forming a primary screening result;
the true value extraction unit is used for extracting farmland/non-farmland true values based on the high-precision farmland boundaries;
the pre-screening unit is used for performing deep learning based on farmland/non-farmland truth values, classifying image blocking areas in the primary screening result, judging whether the current image blocking areas belong to farmland areas, removing the non-farmland areas, and reserving the farmland areas to form a pre-screening result;
and the storage unit is used for storing the pre-screening result of the farmland area in a grid mode.
Embodiment III:
the present invention also provides a memory storing a computer program, the computer program being executable by a processor to:
dividing the satellite image according to grid division to form an image block area;
removing non-farmland areas in the satellite images in advance based on multi-source priori information, and reserving farmland areas to form a preliminary screening result;
extracting a farmland/non-farmland true value based on the high-precision farmland boundary;
deep learning is carried out based on farmland/non-farmland truth values, image blocking areas in the primary screening result are classified, whether the current image blocking areas belong to farmland areas or not is judged, non-farmland areas are removed, farmland areas are reserved, and a pre-screening result is formed;
and storing the pre-screening result of the farmland area in a grid mode.
The invention is preferably realized by using the python programming language under the Ubuntu 16.04 operating system and combining with a TensorFlow machine learning library.
The non-farmland rejection prior information adopted by the invention comprises two types of raster data (such as DEM data, a theme graph and the like) and vector data (such as water distribution); the invention adopts a regression model based on a deep convolutional neural network to realize quick classification of farmland/non-farmland, and can also be realized by selecting other regression models or classification models; the method is mainly used for carrying out large-scale rapid screening on farmland/non-farmland areas, but is also suitable for rapid screening of other ground features (such as forest areas, building areas and the like) after certain adaptation.
Although the present invention has been described in terms of the preferred embodiments, it is not intended to be limited to the embodiments, and any person skilled in the art can make any possible variations and modifications to the technical solution of the present invention by using the methods and technical matters disclosed above without departing from the spirit and scope of the present invention, so any simple modifications, equivalent variations and modifications to the embodiments described above according to the technical matters of the present invention are within the scope of the technical matters of the present invention.
Claims (15)
1. A farmland area pre-screening method based on priori information and deep learning is characterized by comprising the following steps:
dividing the satellite image according to grid division to form an image block area;
removing non-farmland areas in the satellite images in advance based on multi-source priori information, and reserving farmland areas to form a preliminary screening result;
extracting a farmland/non-farmland true value based on the high-precision farmland boundary;
deep learning is carried out based on farmland/non-farmland truth values, image blocking areas in the primary screening result are classified, whether the current image blocking areas belong to farmland areas or not is judged, non-farmland areas are removed, farmland areas are reserved, and a pre-screening result is formed;
storing the pre-screening result of the farmland area in a grid mode;
classifying the image blocking areas by adopting a regression model based on a convolutional neural network, wherein the method specifically comprises the training of the regression model based on deep learning;
the regression model training based on deep learning specifically comprises the following steps:
the satellite image is segmented to form satellite image segmentation blocks;
based on the extracted farmland/non-farmland truth value, calculating a truth value cutting block corresponding to each satellite image cutting block in a farmland/non-farmland truth value diagram;
performing binarization processing on the true value cutting blocks, and calculating the pixel ratio of pixel points in the farmland area in each cutting block to serve as a training label of satellite image cutting blocks;
and inputting the satellite image segmentation blocks and the corresponding training labels into a convolutional neural network for model training.
2. The farmland area pre-screening method based on priori information and deep learning as set forth in claim 1, wherein the pre-removing of non-farmland areas in satellite images specifically comprises the steps of:
performing grid division on the candidate region to obtain grid units;
adopting a theme prior data set as prior information, wherein the theme prior data set comprises a grid data set and a vector data set;
removing grid cells of the non-region of interest based on the land boundary, and reserving the grid cells of the region of interest;
filtering the grid cells based on the grid data set to remove mountain and desert areas;
filtering the grid unit based on the vector data set to remove the water area;
grid cells after removing mountain, desert and water areas are used as candidate farmland areas.
3. The farmland area pre-screening method based on prior information and deep learning of claim 2, wherein the filtering of the grid cells based on the grid data set comprises the steps of:
matching the grid data set with the grid cells by adopting a mapping method to obtain target areas corresponding to the grid cells in the grid data set;
and counting the proportion of the pixels of which the types are mountain or desert in the target area to the pixels of the target area, and judging that the grid unit belongs to the mountain area or the desert area if the proportion is larger than a threshold value, or else, judging that the grid unit belongs to the candidate farmland area.
4. The farmland area pre-screening method based on priori information and deep learning as claimed in claim 2, wherein the specific steps of filtering the grid cells based on the vector data set are as follows:
calculating the area of the grid cells;
calculating the ratio of the sum of the area of the grid unit and the area of the intersection of all the water polygons in the vector data set to the area of the grid unit;
and if the ratio is greater than the threshold value, judging that the grid unit belongs to the water area, otherwise, judging that the grid unit belongs to the candidate farmland area.
5. The farmland area prescreening method based on priori information and deep learning according to claim 1, wherein the farmland/non-farmland truth value is obtained based on high-precision farmland boundary extraction, and a farmland/non-farmland binary mask corresponding to satellite images is generated, specifically comprising the following steps:
selecting satellite images, and extracting farmland boundaries of the satellite images with high precision;
and generating a farmland/non-farmland binary mask by using the farmland boundary, wherein an image area surrounded by the farmland boundary is farmland, and an image area outside the farmland boundary is non-farmland, so as to provide a training truth value for farmland/non-farmland classification based on deep learning.
6. A farmland area pre-screening method based on priori information and deep learning is characterized by comprising the following steps:
dividing the satellite image according to grid division to form an image block area;
removing non-farmland areas in the satellite images in advance based on multi-source priori information, and reserving farmland areas to form a preliminary screening result;
extracting a farmland/non-farmland true value based on the high-precision farmland boundary;
deep learning is carried out based on farmland/non-farmland truth values, image blocking areas in the primary screening result are classified, whether the current image blocking areas belong to farmland areas or not is judged, non-farmland areas are removed, farmland areas are reserved, and a pre-screening result is formed;
storing the pre-screening result of the farmland area in a grid mode;
classifying the image blocking areas by adopting a regression model based on a convolutional neural network, wherein the method specifically comprises quick classification of farmland/non-farmland of the regression model based on deep learning;
the regression model farmland/non-farmland rapid classification based on deep learning specifically comprises the following steps:
cutting the satellite image to form cutting blocks;
inputting each segmentation block into a farmland/non-farmland rapid classification model based on a convolutional neural network, and giving out a farmland duty ratio estimated value of each segmentation block;
and (3) averaging the estimated values of all the cut blocks in the current satellite image, comparing the average value with a threshold value, if the average value is larger than the threshold value, considering that the current satellite image contains farmland areas, and if the average value is smaller than the threshold value, considering that the current satellite image does not contain farmland areas.
7. The farmland area pre-screening method based on priori information and deep learning as set forth in claim 6, wherein the pre-removing of non-farmland areas in satellite images specifically comprises the steps of:
performing grid division on the candidate region to obtain grid units;
adopting a theme prior data set as prior information, wherein the theme prior data set comprises a grid data set and a vector data set;
removing grid cells of the non-region of interest based on the land boundary, and reserving the grid cells of the region of interest;
filtering the grid cells based on the grid data set to remove mountain and desert areas;
filtering the grid unit based on the vector data set to remove the water area;
grid cells after removing mountain, desert and water areas are used as candidate farmland areas.
8. The farmland area pre-screening method based on prior information and deep learning of claim 7, wherein the filtering of the grid cells based on the grid data set comprises the steps of:
matching the grid data set with the grid cells by adopting a mapping method to obtain target areas corresponding to the grid cells in the grid data set;
and counting the proportion of the pixels of which the types are mountain or desert in the target area to the pixels of the target area, and judging that the grid unit belongs to the mountain area or the desert area if the proportion is larger than a threshold value, or else, judging that the grid unit belongs to the candidate farmland area.
9. The farmland area pre-screening method based on priori information and deep learning of claim 7, wherein the specific steps of filtering grid cells based on vector data sets are as follows:
calculating the area of the grid cells;
calculating the ratio of the sum of the area of the grid unit and the area of the intersection of all the water polygons in the vector data set to the area of the grid unit;
and if the ratio is greater than the threshold value, judging that the grid unit belongs to the water area, otherwise, judging that the grid unit belongs to the candidate farmland area.
10. The farmland area prescreening method based on priori information and deep learning according to claim 6, wherein the farmland/non-farmland truth value is obtained based on high-precision farmland boundary extraction, and a farmland/non-farmland binary mask corresponding to satellite images is generated, specifically comprising the following steps:
selecting satellite images, and extracting farmland boundaries of the satellite images with high precision;
and generating a farmland/non-farmland binary mask by using the farmland boundary, wherein an image area surrounded by the farmland boundary is farmland, and an image area outside the farmland boundary is non-farmland, so as to provide a training truth value for farmland/non-farmland classification based on deep learning.
11. A farmland boundary extraction method, characterized by comprising a farmland area prescreening method based on priori information and deep learning according to any one of claims 1-10.
12. Farmland area prescreening device based on priori information and deep learning, characterized in that the device includes:
the image segmentation unit is used for segmenting the satellite image according to grid division to form an image block area;
the primary screening unit is used for removing non-farmland areas in the satellite images in advance based on the multi-source priori information, reserving the farmland areas and forming a primary screening result;
the true value extraction unit is used for extracting farmland/non-farmland true values based on the high-precision farmland boundaries;
the pre-screening unit is used for performing deep learning based on farmland/non-farmland truth values, classifying image blocking areas in the primary screening result, judging whether the current image blocking areas belong to farmland areas, removing the non-farmland areas, and reserving the farmland areas to form a pre-screening result;
the storage unit is used for storing the pre-screening result of the farmland area in a grid mode;
the pre-screening unit classifies the image blocking areas by adopting a regression model based on a convolutional neural network, and specifically comprises regression model training based on deep learning;
the regression model training based on deep learning specifically comprises the following steps:
the satellite image is segmented to form satellite image segmentation blocks;
based on the extracted farmland/non-farmland truth value, calculating a truth value cutting block corresponding to each satellite image cutting block in a farmland/non-farmland truth value diagram;
performing binarization processing on the true value cutting blocks, and calculating the pixel ratio of pixel points in the farmland area in each cutting block to serve as a training label of satellite image cutting blocks;
and inputting the satellite image segmentation blocks and the corresponding training labels into a convolutional neural network for model training.
13. Farmland area prescreening device based on priori information and deep learning, characterized in that the device includes:
the image segmentation unit is used for segmenting the satellite image according to grid division to form an image block area;
the primary screening unit is used for removing non-farmland areas in the satellite images in advance based on the multi-source priori information, reserving the farmland areas and forming a primary screening result;
the true value extraction unit is used for extracting farmland/non-farmland true values based on the high-precision farmland boundaries;
the pre-screening unit is used for performing deep learning based on farmland/non-farmland truth values, classifying image blocking areas in the primary screening result, judging whether the current image blocking areas belong to farmland areas, removing the non-farmland areas, and reserving the farmland areas to form a pre-screening result;
the storage unit is used for storing the pre-screening result of the farmland area in a grid mode;
the pre-screening unit classifies the image blocking areas by adopting a regression model based on a convolutional neural network, and specifically comprises quick classification of a regression model farmland/non-farmland based on deep learning;
the regression model farmland/non-farmland rapid classification based on deep learning specifically comprises the following steps:
cutting the satellite image to form cutting blocks;
inputting each segmentation block into a farmland/non-farmland rapid classification model based on a convolutional neural network, and giving out a farmland duty ratio estimated value of each segmentation block;
and (3) averaging the estimated values of all the cut blocks in the current satellite image, comparing the average value with a threshold value, if the average value is larger than the threshold value, considering that the current satellite image contains farmland areas, and if the average value is smaller than the threshold value, considering that the current satellite image does not contain farmland areas.
14. A memory storing a computer program, wherein the computer program is executed by a processor to:
dividing the satellite image according to grid division to form an image block area;
removing non-farmland areas in the satellite images in advance based on multi-source priori information, and reserving farmland areas to form a preliminary screening result;
extracting a farmland/non-farmland true value based on the high-precision farmland boundary;
deep learning is carried out based on farmland/non-farmland truth values, image blocking areas in the primary screening result are classified, whether the current image blocking areas belong to farmland areas or not is judged, non-farmland areas are removed, farmland areas are reserved, and a pre-screening result is formed;
storing the pre-screening result of the farmland area in a grid mode;
classifying the image blocking areas by adopting a regression model based on a convolutional neural network, wherein the method specifically comprises the training of the regression model based on deep learning;
the regression model training based on deep learning specifically comprises the following steps:
the satellite image is segmented to form satellite image segmentation blocks;
based on the extracted farmland/non-farmland truth value, calculating a truth value cutting block corresponding to each satellite image cutting block in a farmland/non-farmland truth value diagram;
performing binarization processing on the true value cutting blocks, and calculating the pixel ratio of pixel points in the farmland area in each cutting block to serve as a training label of satellite image cutting blocks;
and inputting the satellite image segmentation blocks and the corresponding training labels into a convolutional neural network for model training.
15. A memory storing a computer program, wherein the computer program is executed by a processor to:
dividing the satellite image according to grid division to form an image block area;
removing non-farmland areas in the satellite images in advance based on multi-source priori information, and reserving farmland areas to form a preliminary screening result;
extracting a farmland/non-farmland true value based on the high-precision farmland boundary;
deep learning is carried out based on farmland/non-farmland truth values, image blocking areas in the primary screening result are classified, whether the current image blocking areas belong to farmland areas or not is judged, non-farmland areas are removed, farmland areas are reserved, and a pre-screening result is formed;
storing the pre-screening result of the farmland area in a grid mode;
classifying the image blocking areas by adopting a regression model based on a convolutional neural network, wherein the method specifically comprises quick classification of farmland/non-farmland of the regression model based on deep learning;
the regression model farmland/non-farmland rapid classification based on deep learning specifically comprises the following steps:
cutting the satellite image to form cutting blocks;
inputting each segmentation block into a farmland/non-farmland rapid classification model based on a convolutional neural network, and giving out a farmland duty ratio estimated value of each segmentation block;
and (3) averaging the estimated values of all the cut blocks in the current satellite image, comparing the average value with a threshold value, if the average value is larger than the threshold value, considering that the current satellite image contains farmland areas, and if the average value is smaller than the threshold value, considering that the current satellite image does not contain farmland areas.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811188060.5A CN111028255B (en) | 2018-10-10 | 2018-10-10 | Farmland area pre-screening method and device based on priori information and deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811188060.5A CN111028255B (en) | 2018-10-10 | 2018-10-10 | Farmland area pre-screening method and device based on priori information and deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111028255A CN111028255A (en) | 2020-04-17 |
| CN111028255B true CN111028255B (en) | 2023-07-21 |
Family
ID=70191899
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811188060.5A Active CN111028255B (en) | 2018-10-10 | 2018-10-10 | Farmland area pre-screening method and device based on priori information and deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111028255B (en) |
Families Citing this family (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP7070872B2 (en) * | 2019-08-29 | 2022-05-18 | 株式会社天地人 | Land aptitude evaluation device, land aptitude evaluation method, and program |
| CN112733587A (en) * | 2020-08-10 | 2021-04-30 | 长春理工大学 | Method for monitoring evolution of desertification process |
| CN112102288B (en) * | 2020-09-15 | 2023-11-07 | 应急管理部大数据中心 | Water body identification and water body change detection method, device, equipment and medium |
| WO2022061496A1 (en) * | 2020-09-22 | 2022-03-31 | 深圳市大疆创新科技有限公司 | Object boundary extraction method and apparatus, mobile platform and storage medium |
| CN111950942B (en) * | 2020-10-19 | 2021-01-19 | 平安国际智慧城市科技股份有限公司 | Model-based water pollution risk assessment method and device and computer equipment |
| CN112560929B (en) * | 2020-12-09 | 2022-06-10 | 中国石油大学(华东) | Oil spilling area determining method and device and storage medium |
| CN112580510B (en) * | 2020-12-18 | 2024-05-21 | 广州市城市规划设计有限公司 | Method, device, equipment and storage medium for estimating permeable ground area rate |
| CN112784821A (en) * | 2021-03-06 | 2021-05-11 | 深圳市安比智慧科技有限公司 | Building site behavior safety detection and identification method and system based on YOLOv5 |
| CN116403112B (en) * | 2023-04-11 | 2023-10-27 | 生态环境部卫星环境应用中心 | Method and device for demarcating rural black and odorous water body investigation range |
| CN117665869A (en) * | 2023-11-29 | 2024-03-08 | 武汉大学 | Satellite navigation non-line-of-sight observation detection method based on signal characteristics and machine learning |
| CN117636185B (en) * | 2024-01-26 | 2024-04-09 | 安徽大学 | Pine wood nematode disease detecting system based on image processing |
Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102495412A (en) * | 2011-11-23 | 2012-06-13 | 中国科学院地理科学与资源研究所 | Multi-scale impervious surface information collaborative remote sensing inversion method |
| CN103364781A (en) * | 2012-04-11 | 2013-10-23 | 南京财经大学 | Remote sensing data and geographical information system-based grainfield ground reference point screening method |
| CN103971377A (en) * | 2014-05-27 | 2014-08-06 | 中国科学院遥感与数字地球研究所 | Building extraction method based on prior shape level set segmentation |
| CN105787501A (en) * | 2015-12-17 | 2016-07-20 | 武汉大学 | Vegetation classification method capable of automatically selecting features in power transmission line corridor area |
| CN105957115A (en) * | 2016-05-18 | 2016-09-21 | 河北工程大学 | Crop planting structure remote sensing extraction method under generalized DEM idea |
| CN106250856A (en) * | 2016-08-03 | 2016-12-21 | 武汉大学 | A kind of high-definition picture scene classification method based on non-supervisory feature learning |
| CN106600574A (en) * | 2016-08-25 | 2017-04-26 | 中国科学院遥感与数字地球研究所 | Landslide extraction method based on remote-sensing image and altitude data |
| CN106845343A (en) * | 2016-12-15 | 2017-06-13 | 中国科学院电子学研究所苏州研究院 | A kind of remote sensing image offshore platform automatic testing method |
| CN106897519A (en) * | 2017-02-27 | 2017-06-27 | 中国水利水电科学研究院 | A kind of inland lake gathering ground demarcation method based on DEM |
| CN106991397A (en) * | 2017-03-31 | 2017-07-28 | 中国科学院电子学研究所 | View-based access control model conspicuousness constrains the remote sensing images detection method of depth confidence network |
| CN107563413A (en) * | 2017-08-09 | 2018-01-09 | 千寻位置网络有限公司 | The accurate extracting method of unmanned plane image farmland block object |
| CN108614997A (en) * | 2018-04-04 | 2018-10-02 | 南京信息工程大学 | A kind of remote sensing images recognition methods based on improvement AlexNet |
-
2018
- 2018-10-10 CN CN201811188060.5A patent/CN111028255B/en active Active
Patent Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102495412A (en) * | 2011-11-23 | 2012-06-13 | 中国科学院地理科学与资源研究所 | Multi-scale impervious surface information collaborative remote sensing inversion method |
| CN103364781A (en) * | 2012-04-11 | 2013-10-23 | 南京财经大学 | Remote sensing data and geographical information system-based grainfield ground reference point screening method |
| CN103971377A (en) * | 2014-05-27 | 2014-08-06 | 中国科学院遥感与数字地球研究所 | Building extraction method based on prior shape level set segmentation |
| CN105787501A (en) * | 2015-12-17 | 2016-07-20 | 武汉大学 | Vegetation classification method capable of automatically selecting features in power transmission line corridor area |
| CN105957115A (en) * | 2016-05-18 | 2016-09-21 | 河北工程大学 | Crop planting structure remote sensing extraction method under generalized DEM idea |
| CN106250856A (en) * | 2016-08-03 | 2016-12-21 | 武汉大学 | A kind of high-definition picture scene classification method based on non-supervisory feature learning |
| CN106600574A (en) * | 2016-08-25 | 2017-04-26 | 中国科学院遥感与数字地球研究所 | Landslide extraction method based on remote-sensing image and altitude data |
| CN106845343A (en) * | 2016-12-15 | 2017-06-13 | 中国科学院电子学研究所苏州研究院 | A kind of remote sensing image offshore platform automatic testing method |
| CN106897519A (en) * | 2017-02-27 | 2017-06-27 | 中国水利水电科学研究院 | A kind of inland lake gathering ground demarcation method based on DEM |
| CN106991397A (en) * | 2017-03-31 | 2017-07-28 | 中国科学院电子学研究所 | View-based access control model conspicuousness constrains the remote sensing images detection method of depth confidence network |
| CN107563413A (en) * | 2017-08-09 | 2018-01-09 | 千寻位置网络有限公司 | The accurate extracting method of unmanned plane image farmland block object |
| CN108614997A (en) * | 2018-04-04 | 2018-10-02 | 南京信息工程大学 | A kind of remote sensing images recognition methods based on improvement AlexNet |
Non-Patent Citations (6)
| Title |
|---|
| 冯益明等.基于决策树的戈壁信息提取研究.《干旱区地理》.2013,第125-130页. * |
| 吴波 ; 刘嘉 ; 王宏琦 ; 吴一戎 ; .一种高分辨率遥感图像目标自动提取方法.电子与信息学报.2008,(第11期),全文. * |
| 张友水,冯学智,阮仁宗.基于GIS的BP神经网络遥感影像分类研究.南京大学学报(自然科学版).2003,(第06期),全文. * |
| 潘旭冉等.采用改进全卷积网络的"高分一号"影像居民地提取.《电讯技术》.2018,第119-125页. * |
| 田德宇 ; 张耀南 ; 党辉 ; .深度学习融合多源数据的城市土地利用制图研究.西部大开发(土地开发工程研究).2018,(第06期),全文. * |
| 陈俊任 ; 周晓华 ; 付锋 ; .基于遥感影像的水域覆盖信息提取方法.城市勘测.2013,(第02期),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111028255A (en) | 2020-04-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111028255B (en) | Farmland area pre-screening method and device based on priori information and deep learning | |
| CN108830870B (en) | Satellite image high-precision farmland boundary extraction method based on multi-scale structure learning | |
| Soh et al. | ARKTOS: An intelligent system for SAR sea ice image classification | |
| CN114444791A (en) | Flood disaster remote sensing monitoring and evaluation method based on machine learning | |
| Zylshal et al. | A support vector machine object based image analysis approach on urban green space extraction using Pleiades-1A imagery | |
| CN115452759B (en) | River and lake health index evaluation method and system based on satellite remote sensing data | |
| Sugg et al. | Mapping impervious surfaces using object-oriented classification in a semiarid urban region | |
| CN110569733B (en) | Lake long time sequence continuous water area change reconstruction method based on remote sensing big data platform | |
| Sadeghi et al. | Design and implementation of an expert system for updating thematic maps using satellite imagery (case study: changes of Lake Urmia) | |
| CN113780175B (en) | Remote sensing identification method for typhoon and storm landslide in high vegetation coverage area | |
| Zheng et al. | Single shot multibox detector for urban plantation single tree detection and location with high-resolution remote sensing imagery | |
| Vaidya et al. | Classifying heterogeneous urban form into local climate zones using supervised learning and greedy clustering incorporating Landsat dataset | |
| CN113516060A (en) | Building space fusion discrimination method and device based on multi-temporal high-resolution data | |
| Lobo et al. | Fine-scale mapping of a grassland from digitized aerial photography: An approach using image segmentation and discriminant analysis | |
| CN115858840B (en) | Remote sensing image mosaic method based on scene | |
| CN117392382A (en) | Single tree fruit tree segmentation method and system based on multi-scale dense instance detection | |
| Noppitak et al. | Instance segmentation of water body from aerial image using mask region-based convolutional neural network | |
| CN116206210A (en) | NAS-Swin-based remote sensing image agricultural greenhouse extraction method | |
| CN112580504B (en) | Tree species classification counting method and device based on high-resolution satellite remote sensing image | |
| Sharma et al. | Change detection and feature extraction using high-resolution remote sensing images | |
| Ouchra et al. | Comparison of Machine Learning Methods for Satellite Image Classification: A Case Study of Casablanca Using Landsat Imagery and Google Earth Engine | |
| Zhang et al. | A Mapping Approach for Eucalyptus Plantations Canopy and Single-Tree Using High-Resolution Satellite Images in Liuzhou, China | |
| Evans et al. | A season independent U-net model for robust mapping of solar arrays using Sentinel-2 imagery | |
| CN112949726A (en) | ISCP cloud classification method, system, medium and terminal based on FY-4A satellite | |
| Wang et al. | Information extraction of the vehicle from high-resolution remote sensing image based on convolution neural network |
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 | ||
| CB02 | Change of applicant information |
Address after: 200438 9 / F, 10 / F, 11 / F, 12 / F, 38 Lane 1688, Guoquan North Road, Yangpu District, Shanghai Applicant after: QIANXUN SPATIAL INTELLIGENCE Inc. Address before: Room j165, 1st floor, building 64, 1436 Jungong Road, Yangpu District, Shanghai, 200433 Applicant before: QIANXUN SPATIAL INTELLIGENCE Inc. |
|
| CB02 | Change of applicant information | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |