Remote sensing image supervised classification method based on decision tree correction

Abstract

The invention discloses a remote sensing image supervision and classification method based on decision tree correction, and the method comprises the steps: obtaining a to-be-classified satellite image and a reference image, the reference image being a satellite image which covers a research region of the same sensor and has a verifiable ground feature type; all the images are preprocessed, and image data DN values are converted into earth surface radiation values; cutting the processed image to obtain image data of the area needing to be researched; respectively synthesizing the preprocessed wave band data of each image and various parameter results calculated based on the wave band data into new synthesized image data; drawing a region of interest for the image data of the reference composite image, and making a classification decision tree of the image through machine learning; classifying other to-be-classified composite image data based on the classification decision tree, and drawing new regions of interest for the composite images by referring to the classification result; and utilizing a support vector machine to re-classify the surface features of the synthesized image data. The method improves the classification precision and speed.

Description

Technical Field

[0001] The invention relates to the field of remote sensing image object classification, and in particular to a remote sensing image supervised classification method based on decision tree correction. Background Art

[0002] Since 1972, Landsat satellites have provided a relatively accurate data basis for long-term monitoring and analysis of forest coverage, urban expansion, farmland changes, water resource distribution, etc. The TM sensor carried by Landsat5 has seven bands (wavelength ranges from 0.45 to 2.35 μm, resolution ranges from 30 meters to more than 120 meters). Since it started working on March 1, 1984, it has provided nearly 29 years of earth imaging data. Landsat8 has been in operation since its launch on February 11, 2013, providing higher-resolution image data. However, there are spectral differences, radiometric problems, and geometric correction errors between different models of Landsat series satellites. Therefore, complex correction and processing are often required when comparing results. In addition, the acquisition of Landsat images is limited by weather conditions, such as cloud cover, which may lead to missing or poor quality data in certain areas or certain time periods, thus affecting the accuracy of ground object classification. Furthermore, the sensors of Landsat satellites have limitations in spatial and temporal resolution, and the resolution is low, making it impossible to conduct detailed studies on the types of objects in space.

[0003] Commonly used classification methods include decision tree classification and support vector machine classification. Decision trees display the classification process through a tree structure and can handle data of various types and different measurement scales. They have the advantages of fast establishment speed and small amount of calculation, and are suitable for complex and changeable remote sensing image data. However, the classification effect of this method depends largely on the selection of features, and the relationship between classification accuracy and model complexity needs to be weighed during the classification process. Support vector machine classification performs well in dealing with nonlinear classification problems and can better classify data with complex features. However, this method needs to calculate the distance between each sample point and the decision boundary during the training process, so the training time is long, and a large memory space is required when processing large-scale data sets. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention proposes a supervised classification method for remote sensing images based on decision tree correction. The method identifies different types of land objects by utilizing decision trees, provides a reference for the division of regions of interest, and then combines manual correction and supervised classification to achieve unified standard classification processing of remote sensing images with different resolutions and different numbers of bands in long time series, while significantly improving the classification accuracy and the calculation classification speed.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A supervised classification method for remote sensing images based on decision tree correction includes the following steps:

[0007] S1: Obtain the satellite image to be classified and the reference image, pre-process them, and convert the DN value of the image data into the surface radiation value; crop the processed satellite image to obtain the image data of the required study area; the reference image is a satellite image obtained by the same sensor that covers the study area and the type of the ground object can be verified;

[0008] S2: The band data preprocessed by S1 is combined with various parameter results calculated based on the band data to form new synthetic image data, and the satellite synthetic images to be classified are obtained: R1, R2, ..., R n , and a reference synthetic image R;

[0009] S3: Draw the region of interest for the image data of the reference synthetic image R whose ground object type can be verified, and make a classification decision tree for the image through machine learning;

[0010] S4: Other synthetic image data R1, R2, ..., R to be classified based on the classification decision tree obtained in S3 n Perform classification and draw new regions of interest on these synthetic images based on the classification results;

[0011] S5: Based on the newly drawn regions of interest, the supervised classification of the support vector machine is used to reclassify the objects in these synthetic image data.

[0012] Furthermore, in S2, various parameters calculated based on band data include normalized difference vegetation index NDVI, normalized difference vegetation index EVI, normalized difference water index NDWI, water chlorophyll index Chl a, and water sediment suspended solids index SSC.

[0013] Furthermore, the specific calculation formulas for various parameters calculated based on band data are as follows:

[0014]

[0015]

[0016]

[0017]

[0018]

[0019] SSC0.7874*(Red / Green)-0.1376*(Red / Green)+0.2098

[0020] Among them, Red is the surface reflectivity in the red light band, with a wavelength range of 0.64-0.67μm; Green is the surface reflectivity in the green light band, with a wavelength range of 0.53-0.59μm; Blue is the surface reflectivity in the blue light band, with a wavelength range of 0.45-0.51μm; Nir is the surface reflectivity in the near-infrared band, with a wavelength range of 0.85-0.88μm; Coastal is the surface reflectivity in the coastal band, with a wavelength range of 0.43-0.45μm, and a wavelength range of 0.43-0.45μm; M is a process quantity.

[0021] Furthermore, the S3 specifically includes the following sub-steps:

[0022] S3.1: Obtain the main image features of different land object categories through visual interpretation, field investigation, high-resolution image reference and existing classification results, and determine the type of land object to be classified according to actual needs;

[0023] S3.2: Select the reference synthetic image R, and draw the regions of interest corresponding to different types of landforms according to the spectral characteristics and visual interpretation results of various typical landforms in the study area;

[0024] S3.3: Based on the region of interest drawn in S3.2, perform machine learning on the reference synthetic image R and obtain a decision tree that supports the classification of the image.

[0025] Furthermore, the S4 specifically includes the following sub-steps:

[0026] S4.1: Use the classification decision tree obtained in S3.3 to classify the satellite synthetic images R1, R2, ..., R n Classify each one;

[0027] S4.2: For the satellite synthetic images R1, R2,…, R n For each image in the dataset, the decision tree classification results are compared with the spectral characteristics and texture of the image, the areas where the decision tree misjudged the image are marked, and then the area of ​​interest of each image is remade.

[0028] Furthermore, in S1, the preprocessing of the satellite image to be classified and the reference image includes radiometric calibration, atmospheric correction, and geometric correction.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present invention classifies objects in remote sensing images through synthetic image data, expands the classification condition data source and improves the calculation rate of the classification process. Then, the region of interest is corrected based on the decision tree, which greatly reduces the errors that are easy to occur when classifying medium and low resolution images, can reduce the misjudgment caused by radiation errors caused by solar altitude angle and terrain factors, and reduces the errors caused by subjective visual interpretation of classifiers, thereby improving the accuracy as much as possible. In addition, since the memory and calculation amount required by the support vector machine are large and the difficulty of modifying the sample is large, the decision tree is used to perform pre-classification to provide a reference, and other images are supervised and classified based on the decision tree, which can save the calculation and classification speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the remote sensing image supervised classification method based on decision tree correction of the present invention.

[0032] Figure 2 This is the land classification result map of Hangzhou Bay after reasonable and practical processing by the present invention. DETAILED DESCRIPTION

[0033] The present invention will be described in detail below based on the accompanying drawings and preferred embodiments, and the purpose and effects of the present invention will become more clear. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0034] like Figure 1 As shown, the remote sensing image supervised classification method based on decision tree correction of the present invention comprises the following steps:

[0035] S1: Obtain the satellite images to be classified and the reference images, preprocess them, and convert the DN values ​​of the image data into surface radiation values; crop the processed satellite images to obtain the image data of the required study area.

[0036] One or more satellite image data to be classified are represented by r1, r2,…,r n , the reference image is a satellite image obtained by the same sensor that covers the study area and whose landform type can be verified, denoted as r.

[0037] In this embodiment, the study area is the coastal area of ​​Hangzhou Bay, involving the land and water bodies of Hangzhou Bay. The research objective is to observe the classification of land use types in the Hangzhou Bay area. The research data is Landsat satellite image data from 2015. The satellite images that can be verified by the same sensor covering the study area are Landsat satellite image data from 2021.

[0038] The preprocessing operations specifically include radiation calibration, atmospheric correction, geometric correction, etc. All image preprocessing operations are performed based on ENVI software, and the algorithm used in the atmospheric correction process is the FLAASH atmospheric correction algorithm.

[0039] S2: synthesize the band data preprocessed by S1 and various parameter results calculated based on the band data into new synthetic image data. S2 specifically includes the following sub-steps:

[0040] S2.1: Calculate various index parameter results of each image for auxiliary classification using the surface radiation value of each band processed in S1;

[0041] In this embodiment, since the study area includes land and water, auxiliary classifications such as the normalized vegetation index NDVI, the normalized difference vegetation index EVI, the normalized difference water index NDWI, the water chlorophyll index Chl a, and the water sediment suspended solids index SSC can be calculated. The algorithm formula used is as follows:

[0042] NDVI:

[0043]

[0044] EVI:

[0045]

[0046] NDWI:

[0047]

[0048] Chla:

[0049]

[0050]

[0051] SSC:

[0052] SSC0.7874*(Red / Green)20.1376*(Red / Green)+0.2098

[0053] Among them, Red is the surface reflectivity in the red light band, with a wavelength range of 0.64-0.67μm; Green is the surface reflectivity in the green light band, with a wavelength range of 0.53-0.59μm; Blue is the surface reflectivity in the blue light band, with a wavelength range of 0.45-0.51μm; Nir is the surface reflectivity in the near-infrared band, with a wavelength range of 0.85-0.88μm; Coastal is the surface reflectivity in the coastal band, with a wavelength range of 0.43-0.45μm, and a wavelength range of 0.43-0.45μm; M is a process quantity.

[0054] S2.2: The processed band data in S1 and the various index parameter results obtained in S2.1 are synthesized into new multi-band images, namely, the satellite synthetic images R1, R2, …, R n , and the reference synthetic image R.

[0055] In this embodiment, the pre-processed band data of all the images to be classified and the calculated NDVI, Chl a, and SSC data results are synthesized into new satellite synthetic image data.

[0056] S3: Draw the region of interest for the image data of the reference synthetic image R whose ground object type can be verified, and make a classification decision tree for the image through machine learning; S3 specifically includes the following sub-steps:

[0057] S3.1: Obtain the main image features of different land object categories through visual interpretation, field research, high-resolution image reference and existing classification results, and determine the type of land object to be classified according to actual needs;

[0058] In this embodiment, the types of objects in the research area to be classified are divided into eight categories based on research needs: ocean, fish pond, paddy field, swamp, river, unused land, dry (grass) land and construction land.

[0059] S3.2: Select the reference synthetic image R, and draw the regions of interest corresponding to different types of landforms according to the spectral characteristics and visual interpretation results of various typical landforms in the study area;

[0060] S3.3: Based on the region of interest drawn in S3.2, perform machine learning on the reference synthetic image R and obtain a classification decision tree that supports the classification of the image.

[0061] S4: Classify other synthetic image data to be classified based on the classification decision tree obtained in S3, and draw new regions of interest for these synthetic images with reference to the classification results.

[0062] S4.1: Use the classification decision tree obtained in S3.3 to classify the satellite synthetic images R1, R2, ..., R n Classify each one;

[0063] S4.2: For the satellite synthetic images R1, R2,…, R n For each image in the dataset, the decision tree classification results are compared with the spectral characteristics and texture of the image, the areas where the decision tree misjudged the image are marked, and then the area of ​​interest of each image is remade.

[0064] In this embodiment, the types of land features such as rivers and fish ponds are easily misjudged, and it is necessary to manually redraw the marked misclassified areas into the areas of interest through the decision tree classification results and image texture information, so as to correct the classification results.

[0065] S5: Based on the newly drawn regions of interest, supervised classification is used to reclassify the objects in these synthetic image data.

[0066] Based on the newly drawn region of interest in S4.2, the satellite synthetic images R1, R2, …, R n The ground object type is reclassified to obtain the final classification result, such as Figure 2 shown.

[0067] In order to demonstrate the effect of the method of the present invention, the satellite synthetic images R1, R2, ..., R n The ground feature types are newly classified, and the final classification results are shown in Table 1.

[0068] Table 1 Comparison of classification accuracy results of Hangzhou Bay land

[0069]

[0070] As can be seen from Table 1, in this embodiment, when only the support vector machine is used for classification, the kappa index can only reach about 0.95, and when only the decision tree is used, it can only reach 0.85. After classification by the support vector machine corrected by the decision tree, the accuracy can generally reach above 0.97, which greatly improves the accuracy of classification.

[0071] In addition, when only the decision tree classification method is used, the classification process only screens the spectral value of a single pixel and the band combination results, so the classification speed is fast, but due to the mixed pixel phenomenon in the low- and medium-resolution images, the classification accuracy is low. When only the support vector machine classification method is used, the data is projected to a higher dimension for analysis, and the pixels around the target pixel are extracted to generate new data features, thereby extracting the type information of regular-shaped objects, and the classification accuracy is improved, but due to the large amount of analysis, the calculation speed is slow. The method of the present invention combines the characteristics of the two methods. The region of interest drawn based on the decision tree correction is more reliable, and the speed of reclassification through the support vector machine method can be increased by reducing the number of regions of interest, so that the method of the present invention improves the classification accuracy while accelerating the classification speed.

[0072] Those skilled in the art can understand that the above are only preferred examples of the invention and are not intended to limit the invention. Although the invention is described in detail with reference to the above examples, those skilled in the art can still modify the technical solutions recorded in the above examples or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, etc. made within the spirit and principle of the invention shall be included in the protection scope of the invention.

Claims

1. A supervised classification method for remote sensing images based on decision tree correction, characterized in that: The steps include: S1: Obtain the satellite image to be classified and the reference image, pre-process them, and convert the DN value of the image data into the surface radiation value; crop the processed satellite image to obtain the image data of the required study area; the reference image is a satellite image obtained by the same sensor that covers the study area and the type of the ground object can be verified; S2: The band data preprocessed by S1 is combined with various parameter results calculated based on the band data to form new synthetic image data, and the satellite synthetic images to be classified are obtained: R1, R2, ..., R n , and a reference synthetic image R; S3: Draw the region of interest for the image data of the reference synthetic image R whose ground object type can be verified, and make a classification decision tree for the image through machine learning; S4: Other synthetic image data R1, R2, ..., R to be classified based on the classification decision tree obtained in S3 n Perform classification and draw new regions of interest on these synthetic images based on the classification results; S5: Based on the newly drawn regions of interest, the supervised classification of the support vector machine is used to reclassify the objects in these synthetic image data.

2. The remote sensing image supervised classification method based on decision tree correction according to claim 1 is characterized in that: In S2, various parameters calculated based on band data include normalized difference vegetation index NDVI, normalized difference vegetation index EVI, normalized difference water index NDWI, water chlorophyll index Chl a, and water sediment suspended solids index SSC.

3. The remote sensing image supervised classification method based on decision tree correction according to claim 2 is characterized in that: The specific calculation formulas for various parameters calculated based on band data are as follows: SSC=0.7874*(Red / Green) 2 -0.1376*(Red / Green)+0.2098 Among them, Red is the surface reflectivity in the red light band, with a wavelength range of 0.64-0.67μm; Green is the surface reflectivity in the green light band, with a wavelength range of 0.53-0.59μm; Blue is the surface reflectivity in the blue light band, with a wavelength range of 0.45-0.51μm; Nir is the surface reflectivity in the near-infrared band, with a wavelength range of 0.85-0.88μm; Coastal is the surface reflectivity in the coastal band, with a wavelength range of 0.43-0.45μm, and a wavelength range of 0.43-0.45μm; M is a process quantity.

4. The remote sensing image supervised classification method based on decision tree correction according to claim 2 is characterized in that: The S3 specifically includes the following sub-steps: S3.1: Obtain the main image features of different land object categories through visual interpretation, field investigation, high-resolution image reference and existing classification results, and determine the type of land object to be classified according to actual needs; S3.2: Select the reference synthetic image R, and draw the regions of interest corresponding to different types of landforms according to the spectral characteristics and visual interpretation results of various typical landforms in the study area; S3.3: Based on the region of interest drawn in S3.2, perform machine learning on the reference synthetic image R and obtain a decision tree that supports the classification of the image.

5. The remote sensing image supervised classification method based on decision tree correction according to claim 4 is characterized in that: The S4 specifically includes the following sub-steps: S4.1: Use the classification decision tree obtained in S3.3 to classify the satellite synthetic images R1, R2, ..., R n Classify each one; S4.2: For the satellite synthetic images R1, R2,…, R n For each image in the dataset, the decision tree classification results are compared with the spectral characteristics and texture of the image, the areas where the decision tree misjudged the image are marked, and then the area of ​​interest of each image is remade.

6. The remote sensing image supervised classification method based on decision tree correction according to claim 1 is characterized in that: In S1, the preprocessing of the satellite image to be classified and the reference image includes radiometric calibration, atmospheric correction, and geometric correction.

Images