A method for making a sentinel-2 satellite image dataset for segmentation algorithm

Abstract

The application relates to a Sentinel-2 satellite image data set manufacturing method for a segmentation algorithm, which solves the defect that it is difficult to centrally process satellite images compared with the prior art. The application comprises the following steps: Sentinel image acquisition and pretreatment; normalization processing on the resolution of the pretreated image; image color processing; optical image data set generation; CDL image acquisition; soybean planting area image extraction and binarization processing; and data set label image generation. The application forms a soybean planting area data set with obvious characteristics through normalization, color processing and image cropping on the satellite image.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of satellite image processing, in particular to a Sentinel-2 satellite image dataset production method for segmentation algorithm. BACKGROUND

[0002] High-resolution remote sensing satellite image segmentation is one of the core technologies in remote sensing technology application, and has many practical application basis and premise. How to quickly and accurately segment large-area planting area images is a hot issue in this field.

[0003] At present, after the application of the image segmentation model based on the neural network in the remote sensing satellite image, the preparation of the planting area segmentation training set based on the Sentinel-2 remote sensing image has become a major problem. In particular, some neural network models have special requirements for data sets, such as the requirement of 256*256 pixels for image size. Therefore, how to produce a representative segmentation data set has become an important work.

[0004] The classical method of producing such data sets mainly uses remote sensing satellite images for cutting, and then corresponds to the cutting label image, and manually labels the corresponding one, which is time-consuming and laborious and inaccurate. Especially in the aspect of data labeling, the labeling of the traditional data set relies on manual labeling, and the labeling accuracy is affected by the quality and fatigue of the labeling personnel. How to efficiently and accurately produce training data sets is a core problem. Especially when the data is huge, how to cut and label is a major problem. SUMMARY

[0005] The purpose of the present application is to solve the problem that it is difficult to centrally process satellite images in the prior art, and to provide a Sentinel-2 satellite image dataset production method for segmentation algorithm to solve the above problems.

[0006] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0007] A Sentinel-2 satellite image dataset production method for segmentation algorithm, comprising the following steps:

[0008] 11) Acquisition and preprocessing of Sentinel image: acquire Sentinel-2 L1C multi-spectral data, read the Sentinel-2 L1C multi-spectral data using Sen2cor software, perform atmospheric correction, and generate Sentinel-2 L2A multi-spectral data;

[0009] 12) Normalization processing is performed on the resolution of the preprocessed image;

[0010] 13) image color processing: image color processing is performed on the result of the normalization processing;

[0011] 14) generation of optical image dataset: for one data image of the dataset sample, the image size is obtained, and the gdal module is used to process the raster data, and the data image is cropped with a repetition rate of 0 using the python language, and is normalized to an image set of 256*256 pixels and the coordinate information is retained; all images of the dataset sample are processed by the gdal module to form a square data set, i.e. an optical image dataset;

[0012] 15) obtaining CDL image: selecting UTM coordinate system, obtaining the corresponding Cropland Data Layer image data according to the time and longitude and latitude of the image in the normalization processing of step 12);

[0013] 16) extracting soybean planting area image and performing binaryzation processing: extracting the soybean planting area image from the CDL image, and performing binaryzation processing on the extracted soybean planting area image;

[0014] 17) generation of dataset label image: normalizing the processed label image to 256*256 pixels and retaining the coordinate information to generate a dataset label image.

[0015] The normalization processing of the resolution of the preprocessed image includes the following steps:

[0016] 21) inputting Sentinel-2 L2A level multi-spectral data into Snap software, wherein the Sentinel-2 L2A level multi-spectral data is composed of 13 bands from visible light to short wave infrared, and the band resolution is 60m, 20m and 10m;

[0017] 22) using the resampling function in Snap software to upsample the bands with spatial resolution of 60m and 20m to 10m, thereby realizing image resolution normalization;

[0018] 23) using the format conversion function in Snap software to convert the image with normalized resolution to envi format;

[0019] 24) integrating all envi format images to form a data set.

[0020] The image color processing includes the following steps:

[0021] 31) importing the envi format Sentinel-2 image data in the data set into Envi5.3 software;

[0022] 32) Using the layer stacking function in Envi5.3 software, selecting B4(Red), B3(Green), B2(Blue) bands for band synthesis, selecting UTM coordinate system, generating true color three-band image data;

[0023] 33) All Sentinel-2 image data are subjected to layer stacking operation to form a set of true color three-band image data;

[0024] 34) Coarse cropping of the generated true color three-band image data set, removing images with deep cloud cover area, retaining part of the shallow cloud cover area image, obtaining the research area image as a data set sample.

[0025] The extraction of soybean planting area image and the binaryzation processing include the following steps:

[0026] 41) Import CDL data into arcmap10.8, select soybean planting area attribute set in the attribute table, and convert the raster data set to Shp data format using conversion tools;

[0027] 42) Using Spatial Analyst Tools to extract soybean planting area image from CDL Shp format data;

[0028] 43) Load the soybean planting area image;

[0029] 44) Reclassification function is used to assign 0 to the background of soybean planting area image, i.e. non-soybean planting area, and 1 to soybean planting area, to generate a binary label image;

[0030] 45) All soybean planting area images are subjected to reclassification processing to form a set of binary label images.

[0031] The generation of the data set label image includes the following steps:

[0032] 51) For the binary label image set, obtain a label data image of the sample set, and obtain the image size;

[0033] 52) Import gdal module, use python language, and perform sliding cropping with a repetition rate of 0 on the label image to normalize it to a 256*256 pixel label image set and retain the coordinate information;

[0034] 53) Import all label data images in the sample set into the gdal module to form a square label data set, which is the data set label image.

[0035] Beneficial effects

[0036] The application discloses a Sentinel-2 satellite image data set preparation method for a segmentation algorithm.

[0037] The method realizes semi-automatic preparation of a large data set, and solves the problem of large workload of manual data set labeling when the data volume is large.

[0038] 1. Time efficiency and cost effectiveness: Semi-automatic data set preparation can greatly shorten the time of data set preparation, especially when dealing with a large amount of data, using automated tools can save a lot of time and labor cost.

[0039] 2. Precision and consistency: Automated tools can provide more accurate and consistent labeling results, thereby improving the quality and accuracy of the data set. In addition, automated tools can avoid errors caused by human fatigue and mental state changes.

[0040] 3. Repeatability: Semi-automatic data set preparation can provide repeatability, even on different data sets at different times, using the same automated tools can get similar labeling results.

[0041] 4. Scalability: Automated tools can be easily extended to new data sets and tasks, thereby improving the efficiency and flexibility of data set preparation. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 The method sequence diagram of the application;

[0043] Figure 2 The remote sensing image data set resolution normalization flowchart of the application;

[0044] Figure 3 The image color processing flowchart of the application;

[0045] Figure 4 The image data specification flowchart of the application;

[0046] Figure 5 The normalized optical image sample diagram of the application;

[0047] Figure 6 The label image binarization flowchart of the application;

[0048] Figure 7 The normalized label image sample diagram of the application. DETAILED DESCRIPTION

[0049] In order to make the structural features and the achieved effects of the present application have further understanding and recognition, the following detailed description is given in combination with preferred embodiments and the accompanying drawings.

[0050] As shown in Figure 1 , the method for manufacturing a Sentinel-2 satellite image data set for a segmentation algorithm comprises the following steps:

[0051] Firstly, Sentinel image acquisition and preprocessing: Sentinel-2 L1C multi-spectral data is acquired, Sen2cor software is used to read the Sentinel-2 L1C multi-spectral data, atmospheric correction is performed, and Sentinel-2 L2A multi-spectral data is generated.

[0052] Secondly, the resolution of the preprocessed image is normalized, as shown in Figure 2 , which shows the specific process of multi-spectral data normalization output envi data, wherein bicubic interpolation method is selected in the resampling process, which has high calculation cost but can ensure that the image quality loss after processing is the least.

[0053] (1) The Sentinel-2 L2A multi-spectral data is input into Snap software, wherein the Sentinel-2 L2A multi-spectral data is composed of 13 bands from visible light to short-wave infrared, and the band resolution is 60m, 20m and 10m;

[0054] (2) The resampling function in Snap software is used to upsample the bands with spatial resolution of 60m and 20m to 10m, so as to realize image resolution normalization;

[0055] (3) The format conversion function in Snap software is used to convert the image with normalized resolution into envi format;

[0056] (4) All envi format images are integrated to form a data set.

[0057] Thirdly, image color processing: the normalized result is subjected to image color processing. As shown in Figure 3 , which shows the specific process of envi data generating true color three-band image data, this step ensures that the coordinate systems in each true color image are consistent, and the purpose is to correctly position the cultivated land data layer image in subsequent operations.

[0058] (1) The envi format Sentinel-2 image data in the data set is imported into Envi5.3 software;

[0059] (2) Using the layer stacking function in Envi5.3 software, selecting B4(Red), B3(Green), B2(Blue) bands for band synthesis, selecting UTM coordinate system, generating true color three-band image data;

[0060] (3) All Sentinel-2 image data are subjected to layer stacking operation to form a true color three-band image data set;

[0061] (4) Roughly crop the true color three-band images in the generated data set, remove images with deep cloud cover area, retain part of the shallow cloud cover area image, and obtain the research area image as a data set sample.

[0062] Fourth step, generation of optical image data set: obtain the image size of a data image of the data set sample, import the gdal module to process raster data, use python language to perform sliding cropping with a repetition rate of 0 on the data image, normalize it to an image set of 256*256 pixels and retain the coordinate information; all images in the data set sample are subjected to gdal module processing to form a square data set, i.e. an optical image data set, and the specific process is shown in Figure 4 . The optical image in the generated image data set is shown in Figure 5 , which is composed of 256x256 pixels, each of which contains three-band values to determine the color of the point.

[0063] Step 5, obtain CDL image: select UTM coordinate system, and obtain the corresponding Cropland Data Layer image data according to the time and latitude and longitude of the image image normalized in step 2.

[0064] Step 6, extract soybean planting area image and perform binaryzation processing: extract the soybean planting area image from the CDL image, and perform binaryzation processing on the extracted soybean planting area image. Figure 6 The specific process of generating binaryzation label data from soybean shp data is shown, which needs to ensure that the coordinate system of the soybean planting area image and the true color image is consistent to prevent image misalignment.

[0065] (1) Import CDL data into arcmap10.8, select the soybean planting area attribute set in the attribute table, and use conversion tools to convert the raster data set to Shp data format;

[0066] (2) Use Spatial Analyst Tools to extract the soybean planting area image from the CDL using Shp format data;

[0067] (3) Load the soybean planting area image;

[0068] (4) Use the reclassification function to assign 0 to the background of the soybean planting area image, i.e., the non-soybean planting area, and 1 to the soybean planting area, to generate a binary label image;

[0069] (5) Reclassify all soybean planting area images to form a binary label image set. Figure 7 The binary label data of the spatial distribution of the soybean planting area is shown by Figure 7 It can be seen that the image contains 256x256 pixels, and the pixel value of 1 represents the soybean planting area, which appears white; the pixel value of 0 represents the non-soybean planting area, which appears black.

[0070] Step 7: Generation of dataset label image: normalize the processed label image to 256*256 pixels and keep the coordinate information to generate the dataset label image.

[0071] (1) For the binary label image set, obtain a label data image of the sample set, and obtain the image size;

[0072] (2) Import the gdal module, use python language to perform sliding cutting with a repetition rate of 0 on the label image, normalize it to a 256*256 pixel label image set and keep the coordinate information;

[0073] (3) Import all label data images in the sample set into the gdal module to form a square label data set, which is the dataset label image.

[0074] The above shows and describes the basic principles, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements fall within the scope of the claimed present application. The scope of protection claimed by the present application is defined by the appended claims and their equivalents.

Claims

1. A method for preparing a Sentinel-2 satellite image dataset for a segmentation algorithm, characterized in that: The following steps are involved: 11) Acquisition and preprocessing of Sentinel images: Acquire Sentinel-2L1C multispectral data, read the Sentinel-2L1C multispectral data using Sen2cor software, perform atmospheric correction, and generate Sentinel-2L2A multispectral data; 12) Normalizing the resolution of the preprocessed image; 13) Image color processing: performing image color processing on the result of normalization processing; 14) Generation of the optical image dataset: For one data image in the dataset sample, obtain the image size and import the GDAL module to process the raster data. Using Python, perform sliding cropping on the data image with a repetition rate of 0, normalize it to an image set of 256*256 pixels, and retain the coordinate information. All images in the dataset sample are processed with the GDAL module to form a square data set, i.e., the optical image dataset. 15) Obtain CDL image: Select the UTM coordinate system and obtain the corresponding Cropland Data Layer image data based on the time and longitude and latitude of the image normalized in step 12); 16) Extracting a soybean planting area image and performing binarization processing: extracting a soybean planting area image from the CDL image, and performing binarization processing on the extracted soybean planting area image; 17) Generation of dataset label images: Normalize the processed label images to 256*256 pixels and retain the coordinate information to generate dataset label images.

2. The method for producing a Sentinel-2 satellite image dataset for a segmentation algorithm according to claim 1, characterized in that: The normalization process for the pre-processed image resolution comprises the following steps: 21) Send the Sentinel-2L2A multispectral data into the Snap software. The Sentinel-2L2A multispectral data consists of 13 bands from visible light to shortwave infrared, with band resolutions of 60m, 20m, and 10m. 22) Use the resampling function in the Snap software to upsample the 60m and 20m spatial resolution bands to 10m to achieve image resolution normalization; 23) Use the format conversion function in Snap software to convert the resolution-normalized image into ENVI format; 24) Integrate all images in ENVI format to form a data set.

3. The method for producing a Sentinel-2 satellite image dataset for a segmentation algorithm according to claim 1, characterized in that: The image color processing comprises the following steps: 31) Import the Sentinel-2 image data in ENVI format in the data set into Envi5.3 software; 32) Use the layer stacking function in Envi5.3 software to select the B4Red, B3Green, and B2Blue bands for band synthesis, select the UTM coordinate system, and generate true color three-band image data; 33) All Sentinel-2 image data are layered to form a true color three-band image data set; 34) Roughly crop the true color three-band images in the generated data set, remove the images of the area covered by deep clouds, retain some images of the area covered by shallow clouds, and obtain the images of the study area as the data set samples.

4. The method for producing a Sentinel-2 satellite image dataset for a segmentation algorithm according to claim 1, wherein: The method of extracting the soybean planting area image and performing binarization processing comprises the following steps: 41) Import the CDL data into ArcMap 10.8, select the soybean planting area attribute set in the attribute table, and use the conversion tools to convert the raster data set into the Shp data format; 42) Use Spatial Analyst Tools to extract soybean planting area images from CDL using Shp format data; 43) Loading soybean planting area image; 44) Use the reclassification function to assign a value of 0 to the background of the soybean planting area image, i.e., the non-soybean planting area, and a value of 1 to the soybean planting area to generate a binary label image; 45) All soybean planting area images are reclassified to form a set of binary label images.

5. The method for producing a Sentinel-2 satellite image dataset for a segmentation algorithm according to claim 1, wherein: The generation of the dataset label image includes the following steps: 51) For the binary label image set, obtain a label data image of the sample set and obtain the image size; 52) Import the GDAL module and use Python to perform sliding cropping on the label image with a repetition rate of 0, normalize it into a 256*256 pixel label image set and retain the coordinate information; 53) All label data images in the sample set are imported into the gdal module to form a square label data set, which is the dataset label image.

Images