A method for extracting high-resolution farmland images

Abstract

A high-resolution farmland image extraction method includes (1) merging existing medium spatial resolution farmland remote sensing data products as the initial farmland spatial range, optimizing the products by integrating vegetation index and harmonic time series analysis factors, and removing significant non-farmland areas; (2) using the results of the previous step as a sample library, constructing a spatial-spectral-temporal feature set based on the time series medium-resolution image, and re-extracting farmland through the random forest algorithm; (3) performing multi-scale segmentation on the high spatial resolution image, obtaining segmented vector patches, nesting the results of the previous step into the vector patches, extracting the patch mean and perimeter-area ratio, and processing them using the natural breakpoint method to obtain high-confidence high spatial resolution samples; (4) inputting the high-resolution samples obtained by migration into a semantic segmentation model to extract farmland. This method utilizes existing remote sensing products as prior information, effectively removes erroneous information in the products, reduces the cost of manual sample annotation, and has a high degree of automation.

Description

Technical Field

[0001] This invention relates to the field of remote sensing information extraction technology, and in particular to a method for extracting high-resolution farmland images. Background Technology

[0002] Arable land refers to land used for growing crops, including cultivated land, newly developed, reclaimed, and prepared land, and fallow land (including rotational fallow land and fallow land); land primarily used for growing crops (including vegetables), interspersed with scattered fruit trees, mulberry trees, or other trees; and reclaimed tidal flats and coastal areas that can guarantee one harvest per year on average. Arable land is closely related to economic development and human life. Timely and reliable arable land information is particularly important in crop yield estimation and ecological environment research, and has attracted widespread attention from countries around the world. my country also attaches great importance to the management of arable land. Therefore, obtaining timely and accurate information on arable land distribution is of great significance for arable land protection and sustainable development. In recent years, due to rapid economic development and the rapid advancement of industrialization and urbanization, arable land in some areas has become non-agricultural and non-grain-producing. With the continuous advancement of remote sensing technology, the spatial, spectral, and temporal resolution of remote sensing images is constantly improving. In particular, the free sharing of data from Landsat and Sentinel provides a data foundation for the accurate extraction of arable land information.

[0003] Currently, most remote sensing land cover products containing cultivated land are concentrated at medium or coarse spatial resolution, with fewer high-resolution cultivated land products. Supervised classification is a commonly used and reliable method for extracting land cover information, and sufficient and reliable training samples are crucial for improving its recognition accuracy. In most studies, land cover samples are often collected through visual interpretation and field surveys. However, acquiring sufficient samples over a large scale is usually time-consuming and labor-intensive. Therefore, there is an urgent need to develop methods for automatically acquiring samples to support large-scale cultivated land mapping, especially for research supporting cultivated land recognition in high spatial resolution images. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by providing a high-resolution farmland image extraction method. This method aims to solve the problem of limited identification of farmland products at high spatial resolution in the current remote sensing and mapping field, as well as to reduce the time-consuming and labor-intensive process of acquiring large-scale samples. It provides an automatic sample acquisition method to support large-scale farmland mapping, especially for supporting research on farmland identification in high spatial resolution images.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for extracting high-resolution farmland images includes the following steps:

[0007] S1. Obtain existing 10-meter spatial resolution remote sensing data products, extract images of the area near the farmland, and then take the union of the images to obtain the initial prior farmland range.

[0008] S2. Acquire 10-meter spatial resolution time-series remote sensing images and calculate the time-series normalized vegetation index (NDVI).

[0009] S3. Based on the time-series normalized vegetation index (NDVI) obtained in step S2, calculate its standard deviation (NDVI). std Harmonic timing component coefficients;

[0010] S4. Regarding NDVI in step S3 std The standard deviation and harmonic time-series component coefficients are respectively subjected to Gaussian distribution clustering to obtain the clustered cultivated land area and non-cultivated land area. This is used to correct the prior cultivated land range in step S1 to obtain a 10-meter optimized cultivated land product, which includes the corrected cultivated land area image and the corrected non-cultivated land area image.

[0011] S5. Based on the 10-meter spatial resolution temporal remote sensing image in step S2, extract the texture features, spectral features and temporal features from the image respectively, form a set, and then construct a multi-type image feature set.

[0012] S6. Perform stratified random sampling on the 10-meter optimized farmland product in step S4 to obtain samples of farmland area images and non-farmland area images. Integrate the multi-type image feature sets in step S5 and input both into a random forest classifier to classify and extract the 10-meter reclassified farmland product. This product contains farmland area images and non-farmland area images with multi-type image features.

[0013] S7. Acquire high spatial resolution remote sensing images, and use the 10-meter reclassified farmland products obtained in step S6 as the source sample. Perform multi-scale segmentation on the high spatial resolution remote sensing images to obtain vector patches. Perform overlay analysis on the two to obtain the mean value and perimeter-area ratio of the vector patches respectively. Use the natural breakpoint method to perform threshold segmentation on the indicators to obtain high spatial resolution farmland image samples and non-farmland image samples.

[0014] S8. Input the high spatial resolution farmland samples and non-farmland samples from step S7 into the deep learning semantic segmentation model for training. After training, the model is used to identify the target area and extract high spatial resolution farmland images.

[0015] To optimize the above technical solution, the specific measures also include:

[0016] Furthermore, the 10-meter spatial resolution remote sensing data products used in step S1 are ESAWorldCover and Esri Land Cover.

[0017] Furthermore,

[0018] In step S2, the 10-meter spatial resolution temporal remote sensing imagery used is based on the Sentinel-2 satellite, and all bands except Coastal aerosol, water vapor, and SWIR-Cirrus are used.

[0019] In step S2, the formula for calculating the time-normalized vegetation index (NDVI) is as follows:

[0020]

[0021] In the formula, ρ NIR Represents near-infrared reflectance; ρ R This indicates the reflectivity in the red light band.

[0022] Furthermore, the specific content of step S3 is as follows:

[0023] Standard deviation NDVI std The calculation formula is as follows:

[0024]

[0025] In the formula, NDVI mean N represents the average of multiple NDVI observations; N represents the significant number of multiple NDVI observations; NDVI n This represents the nth NDVI observation;

[0026] Harmonic timing component coefficient a i and b i The calculation formula is as follows:

[0027]

[0028] In the formula, Y NDVI (t j ) represents the time series of NDVI, ε(t) j ) represents the error sequence of NDVI, where t j Indicates the time point at which NDVI is acquired; a0 represents the principal component, which is a constant; i represents the number of periodic terms in the NDVI time series; f represents the frequency; a i and b i It is represented as harmonic timing component coefficient.

[0029] Furthermore, in step S4, the specific content of the prior cultivated land range in the correction step S1 is as follows: using the clustered cultivated land area range as a benchmark, compare the areas in the prior cultivated land range that do not match it, thereby determining the corrected cultivated land area image and the corrected non-cultivated land area image.

[0030] Furthermore, in step S5, the specific content of extracting texture features, spectral features, and temporal features from the image respectively, forming a set, and constructing a multi-type image feature set is as follows:

[0031] The constructed image texture features include variance, contrast, and correlation, which are extracted using the gray-level co-occurrence matrix method.

[0032] The constructed image spectral features include: normalized vegetation index, normalized water index, and the red, green, blue, near-infrared, red-edge, and shortwave infrared bands in the image;

[0033] The constructed time series features include: NDVI standard deviation, NDVI time series harmonic component coefficients, NDVI range, and NDVI minimum value;

[0034] These are used to form a set of multiple types of image features.

[0035] Further, in step S7,

[0036] The spatial resolution of the high spatial resolution remote sensing image is 1 meter.

[0037] The parameters for the multi-scale segmentation algorithm include: scale 100, shape 0.1, and compactness 0.5.

[0038] When performing threshold segmentation, the threshold is determined by the natural breakpoint method. When the average value of the patches is greater than the threshold, it is farmland; when the perimeter-to-area ratio is greater than the threshold, it is field ridge or road.

[0039] Further, in step S8,

[0040] The deep learning semantic segmentation model used is DeepLab V3+;

[0041] Before training, the deep learning semantic segmentation model divides high spatial resolution samples into training, validation, and test sets in a ratio of 5:1:4 for training, validation, and testing.

[0042] The optimizer for the deep learning semantic segmentation model is set to SGD, the batch size is set to 50, the number of training iterations is set to 50, and the learning rate (lr) is set to 10. -3 The loss function is the cross-entropy loss;

[0043] When the deep learning semantic segmentation model is trained and the target region is identified, its batch size is set to 50.

[0044] A computer-readable storage medium storing a computer program that causes a computer to perform the high-resolution farmland image extraction method as described in any of the preceding claims.

[0045] An electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the high-resolution farmland image extraction method as described in any of the preceding claims.

[0046] The beneficial effects of this invention are:

[0047] (1) A solution is proposed to borrow and optimize current arable land products. By making full use of the prior information of existing arable land products, and by using crop time series information and arable land characteristics, erroneous information can be eliminated, thereby improving the quality of arable land products and greatly reducing the cost of manually collecting land cover samples.

[0048] (2) A module for cross-scale migration of samples was designed. Using the results of farmland extraction at medium spatial resolution as the source samples, vector patches were obtained by using the idea of ​​image segmentation to standardize them, and an index was designed to improve their confidence in high spatial resolution images, thereby completing the cross-scale migration of samples.

[0049] (3) This application has achieved good results in the automated extraction of high spatial resolution farmland. Compared with the traditional deep learning scheme for farmland identification based on visual interpretation of samples, it has high operating efficiency and provides a new solution for farmland identification based on high spatial resolution remote sensing images. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the overall technical solution of this application.

[0051] Figure 2 This is a schematic diagram of the regional image and the manually annotated image in the embodiments of this application.

[0052] Figure 3 This is a schematic diagram of the image actually predicted and recognized by the deep learning semantic segmentation model trained in the implementation of this application. Detailed Implementation

[0053] The invention will now be described in further detail with reference to the accompanying drawings.

[0054] This invention focuses on a high spatial resolution remote sensing image acquired from Zibo, Shandong Province. The image has a spatial resolution of 1 meter, includes three RGB bands, and a size of 12951×9462 pixels. The acquisition date was May 12, 2020. Based on this image, corresponding 10-meter spatial resolution remote sensing public products covering the same area, including cultivated land, were acquired: ESAWorldCover 2020 and EsriLand Cover 2020. Simultaneously, Sentinel-2 time-series images covering the entire region for 2020 were downloaded and acquired, also with a spatial resolution of 10 meters.

[0055] The overall flowchart of this implementation case is as follows: Figure 1 As shown, it includes the following steps:

[0056] Step 1: Data Preparation and Preprocessing. A dataset for the Zibo, Shandong study area was acquired, specifically including: remote sensing public products, Sentinel-2 time-series imagery, and high spatial resolution remote sensing imagery. ArcGIS was used to project, register, and crop the images to align the different data in geographic space. Simultaneously, cultivated and non-cultivated land samples in the area were manually labeled for verifying the accuracy of cultivated land identification results. The RGB imagery of this area and its corresponding ground truth surface annotations are shown below. Figure 2 As shown (where (a) is the RGB image of the experimental area and (b) is the ground truth label). In addition, the union of 10-meter spatial resolution remote sensing public goods cultivated land classes was extracted to obtain the initial prior cultivated land range.

[0057] Step 2: Calculate the Time-Normalized Difference Vegetation Index (NDVI). Based on 10-meter resolution Sentinel-2 time-series remote sensing imagery, morphological features are first extracted, and the Time-Normalized Difference Vegetation Index (NDVI) is calculated using the following formula:

[0058]

[0059] In the formula, ρ R ρ is the reflectivity in the red light band. NIR This refers to the reflectivity in the near-infrared band.

[0060] Step 3: Calculate the NDVI standard deviation and NDVI harmonic time-series component coefficients. Based on the time-series NDVI obtained in Step 2, calculate its standard deviation NDVI. std Harmonic timing component coefficient a i and b i .

[0061] Step 4: Gaussian binomial clustering to obtain optimized farmland products. The results obtained in Step 3 are subjected to Gaussian clustering to obtain the clustered farmland and non-farmland areas. Using the clustered farmland area as a benchmark, a mask is used to remove areas that do not match the prior farmland area obtained in Step 1, thereby obtaining 10-meter optimized farmland products.

[0062] Step 5: Construct a multi-type image feature set. Calculate and extract multi-type features from the images separately. Use the gray-level co-occurrence matrix to extract texture features: variance, contrast, and correlation. Construct spectral features of the images: normalized vegetation index, modified normalized water index, and red, green, blue, near-infrared, red edge, shortwave infrared 1, and shortwave infrared 2 bands synthesized from the median and third-quarter spectral density. Construct temporal features including: NDVI standard deviation, NDVI time-series harmonic component coefficients, NDVI extreme value difference, and NDVI minimum value. Based on these, construct a multi-type image feature set.

[0063] Step 6: Reclassify and obtain 10-meter farmland. Stratified random sampling is performed on the 10-meter optimized farmland product from Step 4 to obtain farmland and non-farmland samples. These are then integrated with the multi-type image feature sets from Step 5, and both are input into a random forest classifier. The random forest algorithm parameters are set as follows: the number of classification trees is 500, and the number of features in the subset is set to the square root of the feature count. Finally, the 10-meter reclassified farmland is extracted.

[0064] Step 7: Cross-scale transfer of samples. Using the 10-meter reclassified farmland obtained in Step 6 as the source sample, multi-scale segmentation is performed on the 1-meter spatial resolution remote sensing image to obtain vector patches. The multi-scale segmentation algorithm parameters are set as follows: scale 100, shape 0.1, compactness 0.5. The two images are overlaid and analyzed to obtain two indicators: patch mean and perimeter-area ratio. The natural breakpoint method is used to threshold the indicators. When the patch mean is greater than the threshold identified by the natural breakpoint method, it is farmland; when the perimeter-area ratio is greater than the threshold, it is field ridge or road. These are used to mask the farmland and remove the non-farmland parts, thereby obtaining high spatial resolution farmland samples.

[0065] Step 8: High Spatial Resolution Farmland Extraction. Input the high spatial resolution samples from Step 7 into DeepLabV3+. During model training, the input data consists of labeled 1-meter spatial resolution remote sensing images, with an image size of 128×128. The training, validation, and test sets are split in a 5:1:4 ratio. The optimizer is set to SGD, the batch size to 50, the number of training iterations to 50, and the learning rate (lr) to 10. -3The loss function is CrossEntropy Loss. In the prediction stage, the input image size is 128×128, and the batch size is set to 50. After prediction, the overall accuracy on the test set is 93.80%, the F1-score is 94.16%, and the local details of the recognition results are as follows: Figure 3 As shown, where Figure 3 In the diagram, (a) and (b) are partial actual images of the experimental area, (a1) and (b1) are ground truth labels corresponding to (a) and (b) respectively, i.e., schematic diagrams after manual annotation, and (a2) and (b2) are farmland images identified by the trained deep learning semantic segmentation model corresponding to (a) and (b).

[0066] It should be noted that the terms such as "upper", "lower", "left", "right", "front", and "back" used in the invention are only for clarity of description and are not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention.

[0067] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for extracting high-resolution farmland images, characterized in that, Includes the following steps: S1. Obtain existing 10-meter spatial resolution remote sensing data products, extract images of the area near the farmland, and then take the union of the images to obtain the initial prior farmland range. S2. Acquire 10-meter spatial resolution temporal remote sensing images and calculate the temporal normalized vegetation index. NDVI ; S3. Based on the time-series normalized vegetation index obtained in step S2 NDVI Calculate their standard deviations respectively. NDVI std Harmonic timing component coefficients; the specific content of step S3 is as follows: Standard deviation NDVI std The calculation formula is as follows: ； In the formula, NDVI mean Indicates multiple times NDVI The average of the observed values; N Indicates multiple times NDVI The significant number of observations; NDVI n Indicates the first n Second NDVI Observations; Harmonic timing component coefficients a i and b i The calculation formula is as follows: ； In the formula, Y NDVI ( t j )express NDVI Time series, ( t j )express NDVI The error sequence, where t j Indicates obtaining NDVI Time points; This represents the principal component, which is a constant. i express NDVI The number of periodic terms in a time series; f Indicates frequency; a i and b i Represented as harmonic timing component coefficients; S4. Regarding step S3 NDVI std The standard deviation and harmonic time-series component coefficients are respectively subjected to Gaussian distribution clustering to obtain the clustered cultivated land area and non-cultivated land area. This is used to correct the prior cultivated land range in step S1 to obtain a 10-meter optimized cultivated land product, which includes the corrected cultivated land area image and the corrected non-cultivated land area image. S5. Based on the 10-meter spatial resolution temporal remote sensing image in step S2, extract the texture features, spectral features and temporal features from the image respectively, form a set, and then construct a multi-type image feature set. S6. Perform stratified random sampling on the 10-meter optimized farmland product in step S4 to obtain samples of farmland area images and non-farmland area images. Integrate the multi-type image feature sets in step S5 and input both into a random forest classifier to classify and extract the 10-meter reclassified farmland product. This product contains farmland area images and non-farmland area images with multi-type image features. S7. Acquire high spatial resolution remote sensing images, and use the 10-meter reclassified cultivated land products obtained in step S6 as source samples. Perform multi-scale segmentation on the high spatial resolution remote sensing images to obtain vector patches. Perform overlay analysis on the two images to obtain the mean value and perimeter-area ratio of the vector patches. Use the natural breakpoint method to perform threshold segmentation on the indicators to obtain high spatial resolution cultivated land image samples and non-cultivated land image samples. The spatial resolution of the high spatial resolution remote sensing images is 1 meter. The parameters for the multi-scale segmentation algorithm include: scale 100, shape 0.1, and compactness 0.

5. When performing threshold segmentation, the threshold is determined by the natural breakpoint method. When the average value of the patch is greater than the threshold, it is farmland; when the perimeter-to-area ratio is greater than the threshold, it is field ridge or road. S8. Input the high spatial resolution farmland samples and non-farmland samples from step S7 into the deep learning semantic segmentation model for training. After training, the model is used to identify the target area and extract high spatial resolution farmland images.

2. The method for extracting high-resolution farmland images according to claim 1, characterized in that, The 10-meter spatial resolution remote sensing data products used in step S1 are ESA WorldCover and Esri Land Cover.

3. The method for extracting high-resolution farmland images according to claim 1, characterized in that, In step S2, the 10-meter spatial resolution temporal remote sensing imagery used is based on the Sentinel-2 satellite, and all bands except Coastal aerosol, water vapor, and SWIR-Cirrus are used. In step S2, the time-normalized vegetation index NDVI The calculation formula is as follows: ； In the formula, ρ NIR Indicates near-infrared reflectance; ρ R This indicates the reflectivity in the red light band.

4. The method for extracting high-resolution farmland images according to claim 1, characterized in that, In step S4, the specific content of the prior cultivated land range in the correction step S1 is as follows: using the clustered cultivated land area range as a benchmark, compare the areas in the prior cultivated land range that do not match it, thereby determining the corrected cultivated land area image and the corrected non-cultivated land area image.

5. The method for extracting high-resolution farmland images according to claim 1, characterized in that, In step S5, the specific content of extracting texture features, spectral features, and temporal features from the image, forming a set, and constructing a multi-type image feature set is as follows: The constructed image texture features include variance, contrast, and correlation, which are extracted using the gray-level co-occurrence matrix method. The constructed image spectral features include: normalized vegetation index, normalized water index, and the red, green, blue, near-infrared, red-edge, and shortwave infrared bands in the image; The constructed time series features include: NDVI standard deviation, NDVI time series harmonic component coefficients, NDVI range, and NDVI minimum value; These are used to form a set of multiple types of image features.

6. The method for extracting high-resolution farmland images according to claim 1, characterized in that, In step S8, The deep learning semantic segmentation model used is DeepLab V3+; Before training, the deep learning semantic segmentation model divides high spatial resolution samples into training, validation, and test sets in a ratio of 5:1:4 for training, validation, and testing. The optimizer for the deep learning semantic segmentation model is set to SGD, the batch size is set to 50, the number of training iterations is set to 50, and the learning rate (lr) is set to 10. -3 The loss function is the cross-entropy loss; When the deep learning semantic segmentation model is trained and the target region is identified, its batch size is set to 50.

7. A computer-readable storage medium storing a computer program, characterized in that, The computer program causes the computer to perform the high-resolution farmland image extraction method as described in any one of claims 1-6.

8. An electronic device, characterized in that, include: The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the high-resolution farmland image extraction method as described in any one of claims 1-6.

Images