CN109325451B - High-spatial-resolution cultivated land block full-automatic extraction method based on deep learning - Google Patents
High-spatial-resolution cultivated land block full-automatic extraction method based on deep learning Download PDFInfo
- Publication number
- CN109325451B CN109325451B CN201811118991.8A CN201811118991A CN109325451B CN 109325451 B CN109325451 B CN 109325451B CN 201811118991 A CN201811118991 A CN 201811118991A CN 109325451 B CN109325451 B CN 109325451B
- Authority
- CN
- China
- Prior art keywords
- edge
- extraction
- model
- deep learning
- resolution
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a high-spatial-resolution farmland plot full-automatic extraction technology based on deep learning, which takes the edge extracted by a traditional Canny edge operator as guidance, takes a deep learning theory as a basis, trains an HED deep learning model through a large amount of edge sample label data, and improves parameters such as the network layer number, the pooling size and the like of the model; and further, carrying out cultivated land plot extraction on the images of the research area by using a new network model, and finally carrying out refining and removing treatment on the boundaries extracted by the model by using the boundaries extracted by the Canny edge operator to realize the extraction of the skeleton edges of the cultivated land plots. Compared with the traditional method for manually extracting the plots, the method can effectively improve the production efficiency of the plot extraction and can ensure the uniformity of the edge precision of the plots.
Description
Technical Field
The invention provides a full-automatic cultivated land plot extraction method based on a high-spatial-resolution remote sensing image, which is mainly applied to cultivated land plot extraction based on a high-spatial-resolution remote sensing image and can improve the efficiency of manual plot extraction.
Background
One of the important applications of the remote sensing technology is thematic information extraction, one of the important links in the thematic information extraction process is remote sensing image segmentation, and the image segmentation relates to the extraction of object edge information, so how to rapidly and accurately extract the object edge information is a key step of remote sensing information processing.
The traditional remote sensing image segmentation with better performance mainly adopts a bottom-up aggregation method, the strategy mainly researches the extraction and use of characteristics such as spectrum, texture and shape in the process of area merging, pays attention to the design of aggregation ending conditions and the selection of scale, and algorithms such as FNEA (family planning architecture), watershed and mean shift are used for the generation and merging of initial areas, which also play a key role in the generation and merging of area spectrum characteristics, but because of the complexity of the space form and the distribution structure of ground objects, the generation quality of the initial objects of the remote sensing image is difficult to guarantee, and more fuzzy and space transition areas exist, and the areas are often difficult to be difficult in the merging process; even areas with obvious rules are difficult to ensure that the integrated ground objects are complete and often difficult to correspond. With the development of computer technology, the semantic segmentation theory is gradually formed, and the main idea is to give class information to each pixel, and further realize the segmentation of the image in a convolutional neural network mode, but the method has the problem of low accuracy in the identification of fine ground objects. In addition, many scholars have started to study image segmentation with reference to the object edge, but many studied algorithms focus on local information and do not take into account the overall shape of the object, so that it is difficult to maximize the effect of the algorithm in image segmentation in practical applications.
In view of the fact that edge information has extremely definite indication significance on an image segmentation object, the method takes farmland block extraction as a target, extracts image edges through a large number of high-resolution remote sensing block boundary training deep learning models, takes the image edges as the basis of image top-down division, fuses the bottom gradient calculation results of canny edge operators, forms accurate block boundaries, and accordingly forms complete blocks meeting visual perception.
Disclosure of Invention
In order to solve the current situation of high cost consumption in the existing land block extraction method and fully excavate edge information in an image, the invention provides a high-spatial-resolution remote sensing farmland land block full-automatic extraction method based on edge extraction. The method comprises the following specific steps:
1) collecting and sorting images of a research area, establishing an image library, and simultaneously establishing a cultivated land edge sample library by using a sample label prepared at the early stage;
2) carrying out boundary extraction on the remote sensing image of the research area by using a Canny edge operator;
3) improving the HED model by taking the selected edge label sample data and the corresponding image data as constraints, wherein the parameters comprise the number of network layers and the pooling size;
4) carrying out boundary extraction on the remote sensing image of the target area by using the improved HED model;
5) after the improved model is used for extracting the boundary, refining the edge which is subjected to convolution extraction according to a canny edge operator to obtain the arable land skeleton edge;
6) performing precision inspection on the edge extraction result according to the topological relation and the binary data, returning to the step 1) under the condition that the precision requirement is not met, performing learning model training and farmland plot extraction again according to the reinforcement learning theory,
the method is characterized in that:
performing convolution on the image through a canny edge operator to realize edge extraction, wherein the obtained edge information is used for refining the deep learning model result in the later period; training the model by using the edge label sample data and the corresponding image data, and continuously modifying parameters of the HED model; and (4) refining the HED learning model result by taking the convolution result of the canny edge operator as auxiliary information, and removing the edge information of the non-cultivated land plot.
Compared with the prior image segmentation method, the method has the following characteristics: the method takes the convolution edge result of the traditional canny edge operator as auxiliary information, improves the deep learning model through a large amount of edge label information, further realizes the fine extraction of the cultivated land block, has higher efficiency on the extraction of the cultivated land block, and can save a large amount of cost.
Drawings
FIG. 1 is a flow chart of a method for fully automatically extracting high spatial resolution cultivated land parcels based on deep learning;
FIG. 2A is an original view of a satellite;
FIG. 2B is a diagram illustrating the convolution result of the canny edge operator;
FIG. 3 is a schematic diagram of the HED deep learning model principle;
FIG. 4 is a schematic view of a pooling process;
FIG. 5A is a schematic diagram illustrating the non-extraction result of the cultivated land edge;
and B in the figure 5 is a schematic diagram of arable land edge extraction results.
Detailed Description
Fig. 1 illustrates a main implementation concept of the present invention, wherein key technical parts include automatic retrieval of edge label samples, training of a deep learning model (HED), and edge post-processing with the assistance of Canny edge operator extraction boundary, and the post-processing includes precision verification for edge extraction.
The method comprises the following specific steps:
1) collecting and sorting images of a research area, establishing an image library, and simultaneously establishing a cultivated land edge sample library by using a sample label prepared at the early stage;
2) carrying out boundary extraction on the remote sensing image of the research area by using a Canny edge operator;
3) according to the selected edge label sample data and the corresponding image data as constraints, improving the HED model, including the number of network layers and the pooling size;
4) carrying out boundary extraction on the remote sensing image of the target area by using the improved HED model;
5) after the improved model is used for extracting the boundary, the edge extracted by convolution is refined according to a canny edge operator, and the farmland skeleton edge is obtained.
6) And (3) carrying out precision inspection on the edge extraction result according to the topological relation and the binary data, returning to the step 1) under the condition that the precision requirement is not met, and carrying out learning model training and farmland plot extraction again according to the reinforcement learning theory.
Fig. 2A and 2B illustrate boundary information extracted after a Canny edge convolution operation. Performing convolution on the image through a canny edge operator to realize edge extraction, wherein the obtained edge information is used for refining the deep learning model result in the later period;
FIG. 3 illustrates the HED deep learning model principle. The basic idea of the HED model is edge detection performed end to end, namely, a process from an image to an image, the model emphasizes a process of continuously inheriting and learning in a generated output process to obtain an accurate edge prediction graph, a multi-scale method is mainly applied in a feature learning process, d, e and f in the graph are edge detection results correspondingly obtained by convolutional layers 2, 3 and 4 respectively, and g, h and i are edge extraction performed by a Canny edge operator. The HED model inserts an output layer at the back side of the convolutional layer, and performs supervision on the output layer, so that the result is carried out towards the edge detection direction. And simultaneously, as the backward size of the output layer is smaller, the receptive field is enlarged, and finally, the output under multiple scales is obtained in a weight fusion mode, wherein the receptive field refers to the size of the area of the original image mapped by the pixel points on the characteristic graph output by each layer of the convolutional neural network.
Fig. 4 illustrates a pooling process in deep learning, in which raw data is calculated by pooling of a certain size, and reduced data is extracted. The pooling has the function of compressing the input feature layer, so that the feature map is reduced to simplify the network computation complexity on one hand, and the feature compression and the main feature extraction on the other hand are performed. The pooling process of 2 × 2 is schematically shown in the figure, and 1 to 5 are the result of continuous pooling of 2 × 2.
The drawing A in FIG. 5 shows the specific application of the invention, and the cultivated land plot extraction is carried out in Jiangzhou by applying the method of the invention, wherein white linear information in the drawing A in FIG. 5 is the cultivated land plot boundary automatically extracted through a deep learning model, and the drawing B in FIG. 5 is a result diagram obtained by superposing the extracted boundary information on a remote sensing image, the plot boundary in the result diagram obviously has less extraction omission, the extraction effect is better, and the extraction omission ratio of the whole plot accounts for 15%; the automatic extraction boundary has high accuracy generally, and the modification proportion is lower than 15%.
Through the verification of the actual production test of the Jiangzhou plot, the technical method of the invention improves the efficiency: test forThe mean value of the modification time and efficiency of the two sample sides is 15km2The modification efficiency is expected to be 12km2The hand-drawing of the land parcel is carried out in the early stage, and the efficiency is about 4-7km2The production efficiency is greatly improved.
Therefore, the method can effectively improve the efficiency of the existing cultivated land block extraction method on the premise of ensuring the precision, and can be popularized.
Claims (3)
1. A full-automatic extraction method of high-spatial-resolution cultivated land plots based on deep learning is characterized by comprising the following steps:
1) automatically selecting sample labels from a sample label library according to the image characteristics of a research area;
2) utilizing a Canny edge operator to extract edges;
3) according to the selected edge label sample data and the corresponding image data as constraints, improving the HED model according to the result and the loss condition in the training process, wherein in the improving process of the HED model, the modified parameters comprise the number of network layers and the pooling times;
4) carrying out boundary extraction on the remote sensing image of the target area by using the improved HED model;
5) after the improved model is used for extracting the boundary, refining the edge extracted by the convolutional neural network according to the boundary extracted by the canny edge operator, and obtaining the farmland skeleton edge.
2. The method for fully automatically extracting farmland parcels with high spatial resolution based on deep learning of claim 1, wherein the process of automatically selecting the sample labels from the sample label library is based on the similarity of texture and edge distribution between the samples and the images of the research area.
3. The method for fully automatically extracting high-spatial-resolution farmland plots based on deep learning as claimed in claim 1, wherein in the process of processing after the improved model is used for boundary extraction, the result of the model extraction is refined and eliminated according to the boundary extracted by the earlier Canny edge operator, and skeleton edges are reserved.
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN2017114938586 | 2017-12-31 | ||
CN201711493858 | 2017-12-31 |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109325451A CN109325451A (en) | 2019-02-12 |
CN109325451B true CN109325451B (en) | 2022-05-06 |
Family
ID=65265949
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811118991.8A Active CN109325451B (en) | 2017-12-31 | 2018-09-25 | High-spatial-resolution cultivated land block full-automatic extraction method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109325451B (en) |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110570440A (en) * | 2019-07-19 | 2019-12-13 | 武汉珈和科技有限公司 | Image automatic segmentation method and device based on deep learning edge detection |
CN111259908A (en) * | 2020-03-24 | 2020-06-09 | 中冶赛迪重庆信息技术有限公司 | Machine vision-based steel coil number identification method, system, equipment and storage medium |
CN113361827B (en) * | 2021-07-22 | 2021-11-02 | 四川信息职业技术学院 | Land planning optimization algorithm based on reinforcement learning |
CN113763396B (en) * | 2021-09-02 | 2023-07-25 | 中国农业科学院农业信息研究所 | Remote sensing image land parcel extraction method and system based on deep learning |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101922914A (en) * | 2010-08-27 | 2010-12-22 | 中国林业科学研究院资源信息研究所 | Crown information extraction method and system based on high spatial resolution remote sense image |
CN106384081A (en) * | 2016-08-30 | 2017-02-08 | 水利部水土保持监测中心 | Slope farmland extracting method and system based on high-resolution remote sensing image |
-
2018
- 2018-09-25 CN CN201811118991.8A patent/CN109325451B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101922914A (en) * | 2010-08-27 | 2010-12-22 | 中国林业科学研究院资源信息研究所 | Crown information extraction method and system based on high spatial resolution remote sense image |
CN106384081A (en) * | 2016-08-30 | 2017-02-08 | 水利部水土保持监测中心 | Slope farmland extracting method and system based on high-resolution remote sensing image |
Non-Patent Citations (3)
Title |
---|
基于高分辨率遥感影像的耕地地块提取方法研究;庞新华 等;《测绘科学》;20090120;第34卷(第1期);48-49,161 * |
高分辨率遥感图像耕地地块提取方法研究;胡潭高 等;《光谱学与光谱分析》;20091015;第29卷(第10期);2703-2707 * |
高分辨率遥感土地覆盖分类技术的应用研究--以重庆市黔江贫困区为例;吴田军 等;《地球信息科学》;20160311;第18卷(第3期);353-361 * |
Also Published As
Publication number | Publication date |
---|---|
CN109325451A (en) | 2019-02-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109325451B (en) | High-spatial-resolution cultivated land block full-automatic extraction method based on deep learning | |
CN112287940B (en) | Semantic segmentation method of attention mechanism based on deep learning | |
CN109446992B (en) | Remote sensing image building extraction method and system based on deep learning, storage medium and electronic equipment | |
CN112287807B (en) | Remote sensing image road extraction method based on multi-branch pyramid neural network | |
CN108364288A (en) | Dividing method and device for breast cancer pathological image | |
CN107341517A (en) | The multiple dimensioned wisp detection method of Fusion Features between a kind of level based on deep learning | |
CN107092870A (en) | A kind of high resolution image semantics information extracting method and system | |
CN104517122A (en) | Image target recognition method based on optimized convolution architecture | |
CN114155481A (en) | Method and device for recognizing unstructured field road scene based on semantic segmentation | |
CN106295613A (en) | A kind of unmanned plane target localization method and system | |
CN111080597A (en) | Track fastener defect identification algorithm based on deep learning | |
CN105335965B (en) | Multi-scale self-adaptive decision fusion segmentation method for high-resolution remote sensing image | |
CN105138987A (en) | Vehicle detection method based on aggregation channel characteristic and motion estimation | |
CN113033432A (en) | Remote sensing image residential area extraction method based on progressive supervision | |
CN113269224A (en) | Scene image classification method, system and storage medium | |
CN116630971B (en) | Wheat scab spore segmentation method based on CRF_Resunate++ network | |
Zhu et al. | Change detection based on the combination of improved SegNet neural network and morphology | |
CN111461006A (en) | Optical remote sensing image tower position detection method based on deep migration learning | |
CN105590088A (en) | Traffic sign recognition method based on spare self-encoding and sparse representation | |
CN115546466A (en) | Weak supervision image target positioning method based on multi-scale significant feature fusion | |
CN116883650A (en) | Image-level weak supervision semantic segmentation method based on attention and local stitching | |
CN111611919A (en) | Road scene layout analysis method based on structured learning | |
Han et al. | Multi-stage generative adversarial networks for generating pavement crack images | |
Luo et al. | RBD-Net: robust breakage detection algorithm for industrial leather | |
Ye et al. | FMAM-Net: fusion multi-scale attention mechanism network for building segmentation in remote sensing images |
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 |