A method for identifying agricultural film-covered farmland based on multi-temporal remote sensing images and random forests

Abstract

A method for identifying agricultural film farmland based on multi-temporal remote sensing images and random forests. The present invention relates to a method for identifying agricultural film farmland. The present invention solves the problems of time-consuming and labor-intensive traditional agricultural film monitoring, strong subjectivity, and narrow coverage of field investigations. The method of the present invention: 1. Prepare a classification sample data set; 2. Obtain a multi-band surface reflectance image and calculate 4 common remote sensing index images; 3. Calculate a normalized agricultural film index remote sensing image; 4. Use random forest as a classifier for supervised classification; 5. Evaluate accuracy and generate the final remote sensing classification image. The method of the present invention uses satellite remote sensing technology to overcome the shortcomings of field investigations, has the characteristics of wide coverage, short revisit cycle, low monitoring cost, etc., can realize large-area synchronous observation, and obtain information such as agricultural film coverage area and geographical distribution in real time and accurately, which is of great significance to national agricultural planning, resource management, environmental protection, etc.

Description

Technical Field

[0001] The invention relates to a method for identifying agricultural film and farmland. Background Art

[0002] Film mulching cultivation technology is one of the protective cultivation technologies widely used in the agricultural field. It is one of the important guarantee technologies for the safe supply of agricultural products in China. It has the positive effects of heat preservation and moisture retention, fertilizer and pesticide retention, weed control and salt suppression, promoting crop growth and development, expanding suitable planting areas, and increasing yields. At the same time, it affects the farmland ecological environment, soil microbial community composition and function, soil nutrients, soil structure, greenhouse gas emissions and crop growth. The traditional ground monitoring method of film-covered farmland not only requires rich field work experience and professional knowledge, but is also time-consuming and labor-intensive, highly subjective, and difficult to guarantee the accuracy of interpretation. The field survey has a narrow coverage, a long revisit cycle, and high monitoring costs. Summary of the invention

[0003] In order to solve the problems that traditional agricultural film monitoring is time-consuming, labor-intensive, highly subjective and has a narrow coverage of on-site investigation, the present invention provides an agricultural film farmland identification method based on multi-temporal remote sensing images and random forests.

[0004] The method for identifying agricultural film farmland based on multi-temporal remote sensing images and random forests is carried out in the following steps:

[0005] First, the target area is divided into three types: film-covered farmland, non-film-covered cultivated land and forest land. The three types of classified sample data sets are made using Google Earth high-resolution remote sensing images and ground survey samples, and are randomly divided into training samples and verification samples. The ratio of training samples to verification samples is 7:3;

[0006] 2. After obtaining the three types of original remote sensing images in step 1, the image preprocessing module is used to complete image fusion, radiation calibration, and atmospheric correction to obtain a multi-band surface reflectance image, and calculate the multi-band spectral characteristics and index characteristics; the index characteristics include the Normalized Difference Vegetation Index (NDVI), the Enhanced Vegetation Index (EVI), the Normalized Difference Build-up Index (NDBI), and the Normalized Difference Water Index (NDWI);

[0007] 3. The month with the most obvious difference in the index characteristic time series curve described in step 2 of the comparative analysis is the best time phase for remote sensing information identification. The importance analysis tool is used to calculate the importance scores of different band characteristics to determine the best inversion band. The corresponding calculation equation is selected according to the sensor characteristics to calculate the normalized agricultural film index NAFI. Different calculation formulas are selected according to the sensor characteristics to obtain the agricultural film index remote sensing image. The NAFI corresponding to Landsat and Sentinel are LNAFI and SNAFI respectively.

[0008] Fourth, random forest is used as the classifier for supervised classification. The number of decision trees is set to 300, the number of features remains at the default value, and different combinations are set to use spectral features and index features to participate in image classification;

[0009] 5. Construct a confusion matrix of the verification results, calculate user accuracy, producer accuracy, overall accuracy and Kappa coefficient, evaluate the accuracy and generate the final remote sensing classification image; that is, the identification of farmland covered with agricultural film is realized.

[0010] The present invention proposes a normalized agricultural film index and proposes a specific calculation formula for the characteristics of Sentinel and Landsat sensors. It can distinguish agricultural film from other ground film information more easily than common remote sensing indices (such as NDVI, EVI, etc.), can improve the classification accuracy of random forest classification images, and can be used for fast and accurate agricultural film information extraction. The method of the present invention uses satellite remote sensing technology to overcome the shortcomings of on-site investigations, has the characteristics of wide coverage, short revisit cycle, low monitoring cost, etc., can realize large-area synchronous observation, and obtain information such as agricultural film coverage area and geographical distribution in real time and accurately, which is of great significance to national agricultural planning, resource management, environmental protection, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is the spatial distribution map of classified sample points in 2020 of Example 1;

[0012] Figure 2 This is a time series curve diagram of Sentinel-2 satellite index characteristics in 2020 in Example 1;

[0013] Figure 3 It is the importance of different band characteristics of Sentinel-2 satellite in Example 1;

[0014] Figure 4 Embodiment 1 is an image of Da'an City with different indices based on the Sentinel-2 satellite;

[0015] Figure 5 This is a detailed image of Da'an City with different indexes based on the Sentinel-2 satellite in Example 1;

[0016] Figure 6 The image segmentation results of Da'an City with different index thresholds based on Sentinel-2 satellite in Example 1;

[0017] Figure 7 Example 1 is the spatial distribution and statistical results of film-covered farmland in Da'an City in 2020;

[0018] Figure 8 This is the statistical result of film-covered farmland in various towns and villages in Da'an City in 2020; DETAILED DESCRIPTION

[0019] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0021] Specific implementation method 1: This implementation method of agricultural film farmland identification based on multi-temporal remote sensing images and random forests is carried out according to the following steps:

[0022] First, the target area is divided into three types: film-covered farmland, non-film-covered cultivated land and forest land. The three types of classified sample data sets are made using Google Earth high-resolution remote sensing images and ground survey samples, and are randomly divided into training samples and verification samples. The ratio of training samples to verification samples is 7:3;

[0023] 2. After obtaining the three types of original remote sensing images in step 1, the image preprocessing module is used to complete image fusion, radiation calibration, and atmospheric correction to obtain a multi-band surface reflectance image, and calculate the multi-band spectral characteristics and index characteristics; the index characteristics include Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Normalized Difference Build-up Index (NDBI), and Normalized Difference Water Index (NDWI);

[0024] 3. The month with the most obvious difference in the index characteristic time series curve described in step 2 of the comparative analysis is the best time phase for remote sensing information recognition. The importance analysis tool is used to calculate the importance scores of different band characteristics to determine the best inversion band, and the normalized agricultural film index NAFI is calculated. Different calculation formulas are selected according to the sensor characteristics to obtain the agricultural film index remote sensing image. The NAFI corresponding to Landsat and Sentinel are LNAFI and SNAFI respectively.

[0025] Fourth, random forest is used as the classifier for supervised classification. The number of decision trees is set to 300, the number of features remains at the default value, and different combinations are set to use spectral features and index features to participate in image classification;

[0026] 5. Construct a confusion matrix of the verification results, calculate user accuracy, producer accuracy, overall accuracy and Kappa coefficient, evaluate the accuracy and generate the final remote sensing classification image; that is, the identification of farmland covered with agricultural film is realized.

[0027] In step 1 of this implementation method, the type of film covering in the target area is determined according to the current land use status, distribution range of cultivated land, and ground agricultural conditions survey, and an agricultural film classification system is established, including three types: film-covered farmland, non-filmed cultivated land, and forest land.

[0028] The spectral features and index features used in step 4 of this implementation method need to be selected according to the characteristics of the satellite sensor. For Sentinel-2, 10 spectral features and 5 index features are involved in image classification, where the 10 spectral features include coastal band (Band 1), blue band (Band 2), green band (Band 3), red band (Band 4), red edge band-1 (Band 5), red edge band-2 (Band 6), red edge band-3 (Band 7), near infrared band (Band 8), water vapor band (Band 9), shortwave infrared band-1 (Band 11); the 5 spectral features are NDVI, EVI, NDBI, NDW and SNAFI. For Landsat-8, 7 spectral features and 5 index features are involved in image classification. The 7 spectral features include coastal band (Band 1), blue band (Band 2), green band (Band 3), red band (Band 4), near infrared band (Band 5), shortwave infrared band-1 (Band 6), and shortwave infrared band-2 (Band 7); the 5 spectral features are NDVI, EVI, NDBI, NDW, and LNAFI.

[0029] In step 4 of this implementation, the importance scores of different features are calculated to determine the features involved in the final operation, and the optimal solution is determined by the accuracy index calculated by the confusion matrix and its classification effect.

[0030] Specific implementation method 2: The difference between this implementation method and specific implementation method 1 is that the calculation equations of the index characteristics NDWI, NDVI, NDBI and EVI index in step 2 are respectively:

[0031]

[0032] Among them, NIR is the near infrared band, RED is the red light band, GREEN is the green light band, BLUE is the blue light band, SWIR is the near infrared band, B is the sensitive band for remote sensing identification of film-covered farmland, and t1 and t2 represent the time phases of different remote sensing images. Other steps and parameters are the same as those in the first specific implementation method.

[0033] Specific implementation method three: This implementation method is different from the specific implementation method one in that: the calculation equation of the normalized agricultural film index NAFI in step three is:

[0034]

[0035] Among them, B is the sensitive band for remote sensing identification of film-covered farmland, and t1 and t2 represent the phases of different remote sensing images.

[0036] According to the sensor characteristics of Landsat-8OLI, the Landsat Normalized Agricultural Film Index (LNAFI) is proposed, and its calculation equation is:

[0037]

[0038] Among them, B1 May is the reflectivity of the B1 band of Landsat-8 in May. April This is the reflectivity of the B1 band in April.

[0039] According to the sensor characteristics of Sentinel-2MSI, the Sentinel Normalized Agricultural Film Index (SNAFI) is proposed, and its calculation equation is:

[0040]

[0041] Among them, B2 May is the reflectivity of the B2 band of Sentinel-2 in May. April This is the B2 band reflectivity in April.

[0042] Specific implementation method 4: This implementation method is different from the specific implementation method 1 in that: the accuracy indicators in step 5: user accuracy UA, producer accuracy PA, overall accuracy OA and Kappa coefficient, the calculation equations are:

[0043]

[0044] Where N is the total number of samples; X ii Indicates the number of pixels that are correctly classified; X ij represents the number of samples in the i-th row and j-th column of the error matrix; X ij represents the number of samples in the jth row and ith column of the error matrix; X i+ is the sum of the i-th category in the classification results; X +i is the sum of the i-th category in the ground surface real data; r is the total number of columns in the error matrix; n represents the number of classification categories. Other steps and parameters are the same as those in the first embodiment.

[0045] Example 1 Select Sentinel remote sensing images of Da'an City, Jilin Province on April 6 and May 16, 2020 to determine the remote sensing information extraction process of farmland covered with agricultural film. The specific steps are as follows:

[0046] Step 1: The target area is divided into three types: film-covered farmland, non-film-covered cultivated land, and forest land. Considering the remote sensing interpretation characteristics of film-covered farmland and the film classification system, a classification sample dataset is prepared based on Google Earth high-resolution remote sensing images and ground survey samples. The sample points include 173 film-covered farmlands, 184 non-film-covered cultivated lands, and 184 forest lands. Their spatial distribution is as follows: Figure 1 As shown ( Figure 1 This is the spatial distribution map of the classified sample points in 2020;), randomly divided into training samples and validation samples, with a ratio of 7:3;

[0047] Step 2: After obtaining the original remote sensing image, the image preprocessing module is used to complete image fusion, radiation calibration, and atmospheric correction to obtain a multi-band surface reflectance image and index image time series. Index features include: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Normalized Difference Build-up Index (NDBI), and Normalized Difference Water Index (NDWI); the calculation equations for the index features NDWI, NDVI, NDBI, and EVI are:

[0048]

[0049]

[0050] Among them, NIR is the near infrared band, RED is the red light band, GREEN is the green light band, BLUE is the blue light band, SWIR is the near infrared band, B is the sensitive band for remote sensing identification of mulched farmland, and t1 and t2 represent the phases of different remote sensing images.

[0051] Step 3: Comparative analysis: The month with the most obvious difference in the index feature time series curve described in step 2 is the best time phase for remote sensing information recognition. The importance analysis tool is used to calculate the importance scores of different band features to determine the best inversion band, and the normalized agricultural film index NAFI is calculated. The normalized agricultural film index SNAFI is calculated based on the sensor characteristics of Sentinel. The calculation formula is:

[0052]

[0053] Among them, B2 May is the reflectivity of the B2 band of Sentinel-2 in May. April This is the B2 band reflectivity in April.

[0054] Figure 2 This is a time series curve of the Sentinel-2 satellite index characteristics in 2020. As can be seen from the figure, the different indexes in May and June are significantly different and have distinguishing characteristics. The film covering time in Da'an City was in mid-to-late April, and the appearance of agricultural film significantly changed the surface characteristics. In mid-to-late June, crop growth gradually entered a lush period. The vegetation growth state was the main feature affecting surface changes. The EVI index and NDVI index reached their peaks in July and August. Taking into account the impact of vegetation growth status on remote sensing images in June, it was finally determined that May was the best time phase for remote sensing identification of film-covered farmland, and SNAFI calculations selected images from April and May to complete the calculation. Figure 3 The importance of different band features of the Sentinel-2 satellite in Example 1 is shown, among which the B2 band feature importance score is the highest, and B2 is determined to be the best inversion band. Figure 3 It is the importance of different band characteristics of Sentinel-2 satellite in Example 1;

[0055] Figure 4 Example 1 is based on the different index images of Da'an City from the Sentinel-2 satellite. Figure 5 This is a detailed image of Da'an City with different indexes based on the Sentinel-2 satellite in Example 1. Figure 6 This is the result of the threshold segmentation of different index images of Da'an City based on the Sentinel-2 satellite in Example 1. The red part is the film-covered farmland. By comparing the three pictures, it can be found that SNAFI's ability to distinguish between agricultural film and other ground features is significantly higher than other indices.

[0056] Step 4: Random forest is used as a classifier for supervised classification. The number of decision trees is set to 300, and the number of features remains the default value. The sensor features based on Sentinel-2 use 10 spectral features and 5 index features to participate in image classification. The 10 spectral features include coastal band (Band 1), blue band (Band 2), green band (Band 3), red band (Band 4), red edge band-1 (Band 5), red edge band-2 (Band 6), red edge band-3 (Band 7), near infrared band (Band 8), water vapor band (Band 9), short wave infrared band-1 (Band 11); the 5 spectral features are NDVI, EVI, NDBI, NDW and SNAFI. The accuracy of identifying agricultural film farmland under different feature combination schemes is shown in Table 1. Among them, the S+I combination method includes 10 spectral features and 4 index features; S+I+N includes 10 spectral features and 5 index features, and the SNAFI index is added.

[0057] Table 1 Accuracy recognition results of agricultural film farmland under different feature combination schemes

[0058]

[0059] Step 5: Construct the confusion matrix of the verification results, calculate the user accuracy, producer accuracy, overall accuracy and Kappa coefficient, etc. to evaluate the accuracy and generate the final remote sensing classification image. Among them, the accuracy indicators: user accuracy UA, producer accuracy PA, overall accuracy OA and Kappa coefficient, the calculation equations are:

[0060]

[0061] Where N is the total number of samples; X ii Indicates the number of pixels that are correctly classified; X ij represents the number of samples in the i-th row and j-th column of the error matrix; X ij represents the number of samples in the jth row and ith column of the error matrix; X i+ is the sum of the i-th category in the classification results; X +i is the sum of the i-th category in the real surface data; r is the total number of columns in the error matrix; n represents the number of classification categories.

[0062] Figure 7 It is the spatial distribution and statistical results of film-covered farmland in Da'an City in 2020; Figure 8 This is the statistical result of agricultural land covered with plastic film in various towns in Da'an City in 2020. The total area of ​​agricultural film in Da'an City in 2020 is 31.54km 2 Among them, Sheli Town, Lesheng Town and Honggangzi Township have the largest areas, accounting for 33.5%, 13.5% and 6.2% respectively. Figure 7 and Figure 8 It can be seen that the random forest results considering SNAFI improve the accuracy of mulched farmland producers by 4.76%, the user accuracy by 5.27%, the overall accuracy by 2.49%, and the kappa coefficient by 0.04.

Claims

1. A method for identifying agricultural film fields based on multi-temporal remote sensing images and random forests, characterized in that The method for identifying agricultural film fields based on multi-temporal remote sensing images and random forests is carried out in the following steps: First, the target area is divided into three types: film-covered farmland, non-film-covered cultivated land and forest land. The three types of classified sample data sets are made using Google Earth high-resolution remote sensing images and ground survey samples, and are randomly divided into training samples and verification samples. The ratio of training samples to verification samples is 7:3; 2. After obtaining the three types of original remote sensing images in step 1, the image preprocessing module is used to complete image fusion, radiation calibration, and atmospheric correction to obtain a multi-band surface reflectance image, and calculate the multi-band spectral characteristics and index characteristics and their time series; the index characteristics include the normalized vegetation index NDVI, the enhanced vegetation index EVI, the normalized building index NDBI, and the normalized water index NDWI; 3. The month with the most obvious difference in the index characteristic time series curve described in step 2 of the comparative analysis is the best time phase for remote sensing information recognition. The importance analysis tool is used to calculate the importance scores of different band characteristics to determine the best inversion band, and the normalized agricultural film index NAFI is calculated. Different calculation formulas are selected according to the sensor characteristics to obtain the agricultural film index remote sensing image. The NAFI corresponding to Landsat and Sentinel are LNAFI and SNAFI respectively. The calculation formula of normalized agricultural film index NAFI is: Among them, B is the sensitive band for remote sensing identification of film-covered farmland, t1 and t2 represent the time phases of different remote sensing images; According to the sensor characteristics of Landsat-8OLI, the Landsat Normalized Agricultural Film Index (LNAFI) is proposed, and its calculation equation is: Among them, B1 May is the reflectivity of the B1 band of Landsat-8 in May. April is the reflectivity of the B1 band in April; According to the sensor characteristics of Sentinel-2MSI, the Sentinel Normalized Agricultural Film Index SNAFI is proposed, and its calculation equation is: Among them, B2 May is the reflectivity of the B2 band of Sentinel-2 in May. April is the B2 band reflectivity in April; Fourth, random forest is used as the classifier for supervised classification. The number of decision trees is set to 300, the number of features remains at the default value, and different combinations are set to use spectral features and index features to participate in image classification; 5. Construct a confusion matrix of the verification results, calculate user accuracy, producer accuracy, overall accuracy and Kappa coefficient, evaluate the accuracy and generate the final remote sensing classification image; that is, the identification of farmland covered with agricultural film is realized.

2. The method for identifying agricultural film fields based on multi-temporal remote sensing images and random forests according to claim 1 is characterized in that The index features in step 2 include the normalized vegetation index NDVI, the enhanced vegetation index EVI, the normalized building index NDBI, and the normalized water index NDWI, and their calculation equations are: Among them, NIR is the near infrared band, RED is the red light band, GREEN is the green light band, BLUE is the blue light band, and SWIR is the near infrared band.

3. The method for identifying agricultural film fields based on multi-temporal remote sensing images and random forests according to claim 1 is characterized in that The accuracy indicators in step 5: user accuracy UA, producer accuracy PA, overall accuracy OA and Kappa coefficient, the calculation equations are: Where N is the total number of samples; X ii Indicates the number of pixels that are correctly classified; X ij represents the number of samples in the i-th row and j-th column of the error matrix; X ij represents the number of samples in the jth row and ith column of the error matrix; X i+ is the sum of the i-th category in the classification results; X +i is the sum of the i-th category in the real surface data; r is the total number of columns in the error matrix; n represents the number of classification categories.

Images