CN114463642A - Cultivated land plot extraction method based on deep learning - Google Patents
Cultivated land plot extraction method based on deep learning Download PDFInfo
- Publication number
- CN114463642A CN114463642A CN202111214476.1A CN202111214476A CN114463642A CN 114463642 A CN114463642 A CN 114463642A CN 202111214476 A CN202111214476 A CN 202111214476A CN 114463642 A CN114463642 A CN 114463642A
- Authority
- CN
- China
- Prior art keywords
- model
- deep learning
- label
- land
- parcel
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention provides a farmland plot extraction method based on deep learning, which comprises the following steps: the method comprises the steps of preprocessing remote sensing images, manufacturing land block label samples, establishing a deep learning network model, constructing different land block segmentation network models, performing farmland land block segmentation prediction on images which are not manually labeled by utilizing optimal models, performing post-processing on predicted results, simultaneously using a pseudo label data set and an original label data set obtained by manual labeling as training samples, retraining the models, obtaining a doubled training sample set through pseudo labels, finally obtaining a relatively stable model, and using the model finally obtained by a feature extractor as a prediction model to obtain land block distribution and area. The invention utilizes the deep learning method to improve the precision of the automatic plot segmentation of the computer, reduce the cost of manual identification and greatly save the cost of manpower and time.
Description
Technical Field
The invention relates to the technical field of remote sensing image processing of farmland plot segmentation, in particular to a farmland plot extraction method based on deep learning.
Background
The farmland and the grain safety have close relation, and are important basic resources for human survival and development. The real-time and accurate control of the cultivated land area and distribution is an important scientific basis for agricultural development and regulation, and has wide significance for various applications. Such as dynamic monitoring of arable land, grain productivity surveys, grain safety and agricultural yield forecasts, and the like.
Nowadays, remote sensing images almost cover every place of the global earth surface, and because the remote sensing images have the advantages of wide monitoring range and real-time monitoring, the remote sensing images become important tools for acquiring farmland information and spatial distribution. Most of the traditional image farmland mapping methods are based on machine learning algorithms, such as random forests, support vectors, decision trees and the like. The methods analyze the spectral seasonal characteristics of different land feature types or distinguish different land feature differences by using the characteristics of various indexes, and then the cultivated land extraction is realized by using the spectral seasonal characteristic similarity or index threshold value of the same land feature, but the traditional method can only extract low-grade or medium-grade characteristics from original image data, and has not strong robustness for the cultivated land extraction of a large range and different types of pairs.
Disclosure of Invention
The technical problem to be solved by the invention is to provide a cultivated land plot extraction method based on deep learning aiming at the defects in the prior art.
The technical scheme adopted by the invention for solving the technical problems is as follows:
a farmland plot extraction method based on deep learning comprises the following steps:
step 1: acquiring and preprocessing remote sensing images, namely acquiring high-resolution satellite remote sensing images and Google remote sensing images of cultivated land plots in different season times according to the characteristics of alternate break of cultivated land crops, constructing original image data of the remote sensing images with different resolutions and different season characteristics, and performing preprocessing such as cloud removal and wave band synthesis to obtain processed images;
step 2: making a plot semantic label, and performing vector labeling on the contour of the farmland plot of the image obtained in the step 1 by using an ARCGIS software tool in a manual labeling mode to obtain a vector label graph;
and step 3: performing vector rasterization on the vector label image obtained in the step (2) to obtain a farmland land block semantic label image with the same size as the original image;
and 4, step 4: synchronously cutting the original image obtained in the step 1 and the semantic label graph obtained in the step 3 to obtain a arable land parcel sample library with the size of 512 x 512;
and 5: dividing the sample library obtained in the step 4 into a training sample set, a verification sample set and a test sample set;
step 6: establishing a deep learning network model, and constructing different land parcel segmentation network models by taking a Deeplab V3+ network structure as an overall architecture of a land parcel segmentation model and respectively taking pre-trained Resnet18, Resnet34 and Resnet50 as feature extractors in a coding-decoding mode;
and 7: inputting the training sample set and the verification sample set obtained in the step 5 and the label graph into the network model established in the step 6, and storing the model with the highest precision of each model in the verification samples as an optimal model through multiple iterative training and parameter adjustment;
and 8: and (4) carrying out cultivated land block segmentation prediction on the images which are not manually marked by utilizing the optimal models obtained in the step (7), carrying out model integration on different models by adopting a voting strategy to obtain a prediction result, and carrying out post-processing on the prediction result by adopting hole filling and an edge smoothing algorithm to obtain smoother cultivated land block segmentation pseudo tag data.
And step 9: inputting the pseudo label data set obtained in the step 8 and the original label data set obtained by manual labeling as training samples into the deep learning network model obtained in the step 6 to retrain the model;
step 10: repeating the steps 7-9 for three times, obtaining a doubled training sample set through the pseudo label, and finally obtaining a relatively stable model;
step 11: and (3) taking the model which is finally obtained in the step (10) and takes Resnet50 as a feature extractor as a prediction model, and carrying out land segmentation on the test data.
Step 12: and (4) carrying out grid vectorization on the land parcel segmentation image obtained in the step (11) to obtain the distribution and the area of the land parcel.
Further, preprocessing of radiometric calibration, atmospheric correction, geometric correction, fusion of panchromatic wave bands and multispectral wave bands, cloud removal and wave band combination is carried out on the original satellite remote sensing image data in the step 1.
Further, in the step 1, the resolution of the high-resolution satellite and Google remote sensing images is multi-source data with 0.5 meter and 1 meter, and the land parcels have different seasonal characteristics.
Further, when the plot vector in step 2 is labeled, the value of the farmland plot label is set to 1, and the value of the non-farmland label is set to 0.
Further, in the labeling standard in step 2, the plots distinguishable by human eyes are used as the minimum units and are sketched along the inner edges of the plots, and the inter-plot roads of different plots which can be distinguished obviously need to be distinguished accurately.
Further, step 4 employs a sliding cropping method to crop the original image and the semantic label graph into 512 × 512 pixel samples with a sliding overlap ratio of 10%.
Further, in step 8, in order to improve the quality of the pseudo tag, the prediction model is to perform model integration on different models by using a voting strategy to obtain a prediction result.
Further, in step 9, the pseudo label data set is only added to the training sample, and no verification sample is added to participate in the performance evaluation of the model.
Compared with the prior art, the invention has the following advantages:
1. the invention utilizes the deep learning method to improve the precision of the automatic land segmentation of the computer and reduce the manual identification cost.
2. The method adopts images with various resolutions and seasons as training samples, improves the robustness of the cultivated land parcel recognition model, and reduces the influence of seasonal factors.
3. The invention expands the training sample by manufacturing the pseudo label, and solves the problems of high cost and few samples of manually sketching the sample.
Drawings
FIG. 1 is a flow chart of the working process of the farmland parcel extraction method based on deep learning of the invention.
Detailed Description
The following description of the embodiments of the present invention refers to the accompanying drawings and examples:
it should be noted that the structures, proportions, sizes, and other dimensions shown in the drawings and described in the specification are only for the purpose of understanding and reading the present disclosure, and are not intended to limit the scope of the present disclosure, which is defined by the following claims, and any modifications of the structures, changes in the proportions and adjustments of the sizes, without affecting the efficacy and attainment of the same, are intended to fall within the scope of the present disclosure.
In addition, the terms "upper", "lower", "left", "right", "middle" and "one" used in the present specification are for clarity of description, and are not intended to limit the scope of the present invention, and the relative relationship between the terms and the terms is not to be construed as a scope of the present invention.
Examples
The embodiment of the invention is a farmland plot extraction method based on deep learning, the specific flow refers to fig. 1, and the specific steps of the embodiment of the invention are as follows:
step 1: acquiring and preprocessing remote sensing images, namely acquiring high-resolution satellite remote sensing images and Google remote sensing images of cultivated land plots in different season times according to the characteristics of alternate break of cultivated land crops, constructing original image data of the remote sensing images with different resolutions and different season characteristics, and performing preprocessing such as cloud removal and wave band synthesis to obtain processed images;
step 2: making a plot semantic label, and performing vector labeling on the contour of the farmland plot of the image obtained in the step 1 by using an ARCGIS software tool in a manual labeling mode to obtain a vector label graph;
and step 3: performing vector rasterization on the vector label image obtained in the step (2) to obtain a farmland land block semantic label image with the same size as the original image;
and 4, step 4: synchronously cutting the original image obtained in the step 1 and the semantic label graph obtained in the step 3 to obtain a arable land parcel sample library with the size of 512 x 512;
and 5: dividing the sample library obtained in the step 4 into a training sample set, a verification sample set and a test sample set;
step 6: establishing a deep learning network model, and constructing different land parcel segmentation network models by taking a Deeplab V3+ network structure as an overall architecture of a land parcel segmentation model and respectively taking pre-trained Resnet18, Resnet34 and Resnet50 as feature extractors in a coding-decoding mode;
and 7: inputting the training sample set and the verification sample set obtained in the step 5 and the label graph into the network model established in the step 6, and storing the model with the highest precision of each model in the verification samples as an optimal model through multiple iterative training and parameter adjustment;
and 8: and (4) carrying out cultivated land block segmentation prediction on the image which is not manually marked by utilizing each optimal model obtained in the step (7), carrying out model integration on different models by adopting a voting strategy to obtain a prediction result, and carrying out post-processing on the prediction result by adopting hole filling and an edge smoothing algorithm to obtain smoother cultivated land block segmentation pseudo tag data.
And step 9: inputting the pseudo label data set obtained in the step 8 and the original label data set obtained by manual labeling as training samples into the deep learning network model obtained in the step 6 to retrain the model;
step 10: repeating the steps 7-9 for three times, obtaining a doubled training sample set through the pseudo label, and finally obtaining a relatively stable model;
step 11: and (3) taking the model which is finally obtained in the step (10) and takes Resnet50 as a feature extractor as a prediction model, and carrying out land segmentation on the test data.
Step 12: and (4) carrying out grid vectorization on the land parcel segmentation image obtained in the step (11) to obtain the distribution and the area of the land parcel.
Specifically, the raw satellite remote sensing image data in the step 1 is subjected to radiometric calibration, atmospheric correction, geometric correction, fusion of panchromatic wave bands and multispectral wave bands, cloud removal and wave band combination preprocessing.
Specifically, in the step 1, the resolution of the high-resolution satellite and Google remote sensing images is multi-source data with 0.5 meter and 1 meter, and the land parcels have different seasonal characteristics.
Specifically, when the plot vector is labeled in step 2, the value of the cultivated land plot label is set to 1, and the value of the uncultivated land label is set to 0.
Specifically, in the labeling standard in step 2, the plots distinguishable by human eyes are used as the minimum units and are sketched along the inner edges of the plots, and the inter-plot roads of different plots which can be distinguished obviously need to be distinguished accurately.
Specifically, step 4 employs a sliding cropping method to crop the original image and the semantic label graph into 512 × 512 pixel samples with a sliding overlap ratio of 10%.
Specifically, in step 8, in order to improve the quality of the pseudo tag, the prediction model is to perform model integration on different models by using a voting strategy to obtain a prediction result.
Specifically, in step 9, the pseudo label data set is only added to the training sample, and no verification sample is added to participate in the performance evaluation of the model.
Although the preferred embodiments of the present invention have been described in detail with reference to the accompanying drawings, the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present invention.
Many other changes and modifications can be made without departing from the spirit and scope of the invention. It is to be understood that the invention is not to be limited to the specific embodiments, but only by the scope of the appended claims.
Claims (8)
1. A farmland plot extraction method based on deep learning is characterized in that: the method comprises the following steps:
step 1: acquiring and preprocessing remote sensing images, namely acquiring high-resolution satellite remote sensing images and Google remote sensing images of cultivated land plots in different season times according to the characteristics of alternate break of cultivated land crops, constructing original image data of the remote sensing images with different resolutions and different season characteristics, and performing preprocessing such as cloud removal and wave band synthesis to obtain processed images;
step 2: making a plot semantic label, and performing vector labeling on the contour of the farmland plot of the image obtained in the step 1 by using an ARCGIS software tool in a manual labeling mode to obtain a vector label graph;
and step 3: performing vector rasterization on the vector label image obtained in the step (2) to obtain a farmland plot semantic label image with the same size as the original image;
and 4, step 4: synchronously cutting the original image obtained in the step 1 and the semantic label graph obtained in the step 3 to obtain a arable land parcel sample library with the size of 512 x 512;
and 5: dividing the sample library obtained in the step 4 into a training sample set, a verification sample set and a test sample set;
step 6: establishing a deep learning network model, and constructing different land parcel segmentation network models by taking a Deeplab V3+ network structure as an overall architecture of a land parcel segmentation model and respectively taking pre-trained Resnet18, Resnet34 and Resnet50 as feature extractors in a coding-decoding mode;
and 7: inputting the training sample set and the verification sample set obtained in the step 5 and the label graph into the network model established in the step 6, and storing the model with the highest precision of each model in the verification samples as an optimal model through multiple iterative training and parameter adjustment;
and 8: carrying out cultivated land block segmentation prediction on the image which is not manually marked by utilizing each optimal model obtained in the step (7), carrying out model integration on different models by adopting a voting strategy to obtain a prediction result, and carrying out post-processing on the prediction result by adopting hole filling and an edge smoothing algorithm to obtain smoother cultivated land block segmentation pseudo label data;
and step 9: inputting the pseudo label data set obtained in the step 8 and the original label data set obtained by manual labeling as training samples into the deep learning network model obtained in the step 6 to retrain the model;
step 10: repeating the step 7 to the step 9 for three times, obtaining a doubled training sample set through the pseudo label, and finally obtaining a relatively steady model;
step 11: and (3) taking the model which is finally obtained in the step (10) and takes Resnet50 as a feature extractor as a prediction model, and carrying out land segmentation on the test data.
Step 12: and (4) carrying out grid vectorization on the land parcel segmentation image obtained in the step (11) to obtain the distribution and the area of the land parcel.
2. The cultivated land parcel extraction method based on deep learning of claim 1, characterized by that, the raw high-resolution satellite remote sensing image data in step 1 is preprocessed by radiometric calibration, atmospheric correction, geometric correction, fusion of panchromatic wave band and multispectral wave band, cloud removal and wave band combination.
3. The method for extracting arable land parcels based on deep learning of claim 1, wherein in step 1, the resolution of the high-resolution satellite and Google remote sensing images is 0.5 meter and 1 meter of multi-source data, and the parcels have different seasonal characteristics.
4. The method for extracting cultivated land parcel based on deep learning as claimed in claim 1, characterized in that, in step 2, the cultivated land parcel label value is set to 1 and the non-cultivated land parcel label value is set to 0 when parcel vector is labeled.
5. The method for extracting cultivated land parcels based on deep learning of claim 1, wherein the labeling standard in step 2 is drawn along the inner edge of the parcel with the parcel distinguishable by human eyes as the minimum unit, and the inter-land road surfaces of different parcels which can be distinguished clearly need to be distinguished accurately.
6. The method for extracting arable land parcel based on deep learning of claim 1, characterized in that, step 4 adopts a sliding clipping method, and the original image and semantic tag map are clipped to 512 by 512 pixel sample with 10% sliding overlap rate.
7. The method for extracting farmland plots based on deep learning of claim 1, wherein in step 8, in order to improve the quality of the pseudo labels, the prediction model is a model integration of different models by using a voting strategy to obtain a prediction result.
8. The arable land parcel extraction method based on deep learning of claim 1, characterized in that, in step 9, the pseudo label data set is added to the training sample only, and the performance evaluation of the model is participated in without adding the verification sample.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111214476.1A CN114463642A (en) | 2021-10-19 | 2021-10-19 | Cultivated land plot extraction method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111214476.1A CN114463642A (en) | 2021-10-19 | 2021-10-19 | Cultivated land plot extraction method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114463642A true CN114463642A (en) | 2022-05-10 |
Family
ID=81405875
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111214476.1A Pending CN114463642A (en) | 2021-10-19 | 2021-10-19 | Cultivated land plot extraction method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114463642A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115035422A (en) * | 2022-08-15 | 2022-09-09 | 杭州航天星寰空间技术有限公司 | Data augmentation method and segmentation method for soil planting structure in remote sensing image area |
-
2021
- 2021-10-19 CN CN202111214476.1A patent/CN114463642A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115035422A (en) * | 2022-08-15 | 2022-09-09 | 杭州航天星寰空间技术有限公司 | Data augmentation method and segmentation method for soil planting structure in remote sensing image area |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108573276B (en) | Change detection method based on high-resolution remote sensing image | |
| CN111738165B (en) | Method for extracting individual plant canopy from high-resolution unmanned aerial vehicle visible light remote sensing image | |
| Halme et al. | Utility of hyperspectral compared to multispectral remote sensing data in estimating forest biomass and structure variables in Finnish boreal forest | |
| CN112183209B (en) | Regional crop classification method and system based on multidimensional feature fusion | |
| CN108428220B (en) | Automatic geometric correction method for ocean island reef area of remote sensing image of geostationary orbit satellite sequence | |
| CN109448127B (en) | Farmland high-precision navigation map generation method based on unmanned aerial vehicle remote sensing | |
| CN106897707B (en) | Characteristic image time sequence synthesis method and device based on multi-source midsplit | |
| CN111898503B (en) | Crop identification method and system based on cloud coverage remote sensing image and deep learning | |
| CN111008664B (en) | Hyperspectral sea ice detection method based on space-spectrum combined characteristics | |
| CN113205014B (en) | Time sequence data farmland extraction method based on image sharpening | |
| CN113869370A (en) | Tree species refined classification method based on deep learning algorithm and time sequence sentinel image | |
| CN115880487A (en) | Forest laser point cloud branch and leaf separation method based on deep learning method | |
| CN113252584A (en) | Crop growth detection method and system based on 5G transmission | |
| CN114092794A (en) | Sea ice image classification method, system, medium, equipment and processing terminal | |
| Suarez et al. | Learning image vegetation index through a conditional generative adversarial network | |
| CN114463642A (en) | Cultivated land plot extraction method based on deep learning | |
| CN111832480B (en) | Remote sensing identification method for rape planting area based on spectral characteristics | |
| CN111666914B (en) | Cultivated land identification method, system, equipment and storage medium based on distance between curves | |
| Zou et al. | The fusion of satellite and unmanned aerial vehicle (UAV) imagery for improving classification performance | |
| CN110175638B (en) | Raise dust source monitoring method | |
| Hall | Remote sensing applications in viticulture: recent advances and new opportunities | |
| Yang et al. | Texture feature extraction of mountain economic forest using high spatial resolution remote sensing images | |
| Zhang et al. | A Mapping Approach for Eucalyptus Plantations Canopy and Single-Tree Using High-Resolution Satellite Images in Liuzhou, China | |
| CN114494586A (en) | Lattice projection deep learning network broad-leaved tree branch and leaf separation and skeleton reconstruction method | |
| CN113762128A (en) | Hyperspectral image classification method based on unsupervised learning |
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 |