Planting structure identification system and method based on GEE multiple extraction

Abstract

The invention relates to the technical field of remote sensing classification, in particular to a planting structure recognition system and method based on GEE multiple extraction. Extracting a land utilization type basic element by using the high-resolution image; selecting a gradient boosting tree, a random forest, a decision tree, a support vector machine and a naive Bayesian machine learning model to extract the land utilization type without the basic elements; evaluating the extraction precision of the five machine learning models to obtain a model with the best land utilization type classification effect; and extracting land utilization types of cultivated land attributes, carrying out precision evaluation on the five machine learning models, obtaining a model with the best planting structure classification effect, and identifying the model. According to the method, the problem that the planting structure cannot be accurately identified due to the fact that some different ground features are extremely similar in feature expression is solved, and the method is of great significance for accurately mastering dynamic changes of regional crops, optimizing allocation of resources in the region and comprehensive agricultural development.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing classification, and in particular relates to a planting structure recognition system and method based on GEE (Google Earth Engine) multiple extraction. Background Art

[0002] In order to meet the sustainable development of agricultural resources, it is necessary to timely and accurately grasp the crop types, planting areas and spatial distribution. The extraction of crop planting area and spatial distribution information has important practical significance in crop health monitoring, food yield prediction and national agricultural policy adjustment. Compared with traditional planting structure detection methods, remote sensing monitoring is a broader direction and plays an increasingly important role in precision agriculture. Remote sensing monitoring relies on various satellites. In recent years, Sentinel-2 satellite, as a high-resolution satellite with free access, has been widely used in planting structure and has achieved remarkable results.

[0003] Traditional planting structure extraction methods mainly use spectral indices to extract different land features, such as using the NDVI spectral index to distinguish different land cover conditions. Compared with traditional planting structure recognition methods, machine learning algorithms have become one of the main methods for extracting planting structures after traditional spectral index extraction methods. Machine learning algorithms can quickly and accurately analyze large amounts of remote sensing data to identify different crop types. GEE has become an important tool for planting structure extraction due to its advantages such as large data processing capabilities, rich computing resources, and convenient algorithm application.

[0004] Existing methods usually focus on the overall extraction of planting structures. However, there are some problems in actual situations. Since some different objects have similar features, it is very difficult to distinguish them accurately. To solve this problem, this paper proposes a planting structure recognition system and method based on GEE multiple extraction. Summary of the invention

[0005] In view of the problems existing in the prior art, the object of the present invention is to provide a planting structure recognition system and method based on GEE multiple extraction.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A planting structure recognition system and method based on GEE multiple extraction, comprising the following steps:

[0008] S1. Extract basic elements of land use types using high-resolution images;

[0009] S2. Calculate the NDVI values ​​for different land use types and different planting structures;

[0010] S3. Call Sentinel-2 remote sensing images in GEE and import sample points and boundaries;

[0011] S4, extracting the band reflectance and NDVI value of the sample point as characteristic parameters;

[0012] S5, machine learning model extracts land use types after removing basic elements;

[0013] S6. Compare the accuracy of different models to obtain the optimal land use type extraction model;

[0014] S7, extracting land use type of cultivated land attributes;

[0015] S8, machine learning model extracts the planting structure of cultivated land;

[0016] S9. Compare the accuracy of different models to obtain the optimal implant structure extraction model and identify it.

[0017] Furthermore, in S1, the specific steps of extracting the basic elements of the land use type are as follows:

[0018] S1.1. Obtain the latest Tiandi Map image data as the base map;

[0019] S1.2. Select the basic elements of land use types, create features in ArcGIS and extract them through visual interpretation.

[0020] Furthermore, in S2, the NDVI calculation formula is:

[0021]

[0022] Where NIR is the reflectivity of the B8 near-infrared band; Red is the reflectivity of the B4 red light band.

[0023] Furthermore, in S3, the specific steps for calling remote sensing images are as follows:

[0024] S3.1. Image call: Select L2A-level data that has been orthorectified, geometrically corrected, and atmospherically corrected by Sentinel-2, and call it directly from the GEE cloud platform;

[0025] S3.2. Image selection rules: Screen images with cloud cover less than 10% and crop and synthesize the image data according to the shape of the study area.

[0026] Furthermore, in S5, the main principles of the machine learning model are as follows:

[0027] S5.1. Decision tree is a machine learning algorithm that makes decisions based on a tree structure. It performs a series of feature tests on the data and divides the data into different branches according to the test results, eventually forming a tree structure. Each internal node represents a test of a feature attribute, the branch represents the test result, and the leaf node represents the final classification result. When constructing a decision tree, information gain, information gain rate, Gini index and other indicators are usually used to select the optimal features for division. The indicator principle formula is as follows:

[0028]

[0029]

[0030]

[0031]

[0032]

[0033]

[0034] Where D is the data set; Ent(D) represents the information entropy of the data set D; P k represents the proportion of category k in the data set D, |y| is the number of categories; Gain(D, a) is the information gain; Gain_ratio(D, a) is the information gain rate; Gini(D) is the Gini value; Gini_index(D, a) is the Gini index; a is a certain feature; V is the number of values ​​of feature a; D v is the data subset when feature a takes the value v;

[0035] S5.2. Random forest is a machine learning algorithm that consists of multiple decision trees. When constructing a random forest, multiple sample sets are first randomly extracted from the original training data. Each sample set is used to train a decision tree. When constructing each decision tree, a part of the features is also randomly selected for division. Finally, the final classification result is determined by voting.

[0036] S5.3. The gradient boosting tree model iteratively trains a series of decision trees and combines them to obtain the final prediction result. In each iteration, the algorithm builds a new decision tree based on the prediction error of the current model, so that the new decision tree can reduce the error of the current model. The main principle formula is as follows:

[0037]

[0038] F m+1 (x) = F m (x)+h m(x)

[0039] In the formula, x is the eigenvector; x i is the feature vector of the i-th sample point; F m (x) is the model after m iterations; y i is the true value; L(y i , F m (x i )) is the loss function; r mi is the negative gradient of the loss function, and the training target of the new decision tree is determined by calculating the negative gradient; h m (x) is the newly constructed decision tree;

[0040] S5.4. Support vector machine is a binary classification model. Its basic idea is to find a hyperplane that can separate data points of different categories as correctly as possible and maximize the distance from the point closest to the hyperplane to the hyperplane. For the linearly separable case, the classification decision function is:

[0041] f(x)=sign(w T x+b)

[0042] In the formula, w is the normal vector; b is the intercept; x is the data point;

[0043] For the case of linear inseparability, the kernel function is introduced to map the data to a high-dimensional space, so that the data in the high-dimensional space becomes linearly separable. The classification decision function is:

[0044]

[0045] In the formula, b is the intercept; a i is the Lagrange multiplier; y i is the category label; n n is the sample size; K(x i ,x) is the kernel function;

[0046] S5.5. Naive Bayes is a classification method based on Bayes’ theorem and the conditional independence assumption of features. For given input data, Naive Bayes calculates the probability of each feature appearing in each category, and then calculates the posterior probability that the data belongs to each category according to Bayes’ theorem. Finally, the data is assigned to the category with the largest posterior probability. The main principle formula is as follows:

[0047]

[0048] Where P(C|X) is the posterior probability of belonging to category C given feature X; P(X|C) is the probability of feature X appearing under category C; P(C) is the prior probability of category C; and P(X) is the marginal probability of feature X.

[0049] Furthermore, in S6, the overall accuracy (OA), producer accuracy (PA), user accuracy (UA) and Kappa coefficient are used to compare the accuracy evaluation parameters of different machine learning models. The specific calculation formula is:

[0050]

[0051]

[0052]

[0053]

[0054] In the formula, x is the number of validation samples; and is the total number of samples of the i-th category in the classification results and the total number of samples of the i-th category in the validation samples; X ii is the number in the i-th row and i-th column of the error matrix; it represents the number of correctly classified samples in the i-th category; k is the number of classification categories.

[0055] The beneficial effects of the present invention are:

[0056] The present invention adopts a multiple extraction method combined with different machine learning models, and selects the model with the highest accuracy to extract the land use type and planting structure respectively, thereby ensuring the accuracy of planting structure identification under complex planting conditions;

[0057] The present invention performs step-by-step identification of land use types and planting structures, thereby realizing the extraction of compound rotation planting structures. Compared with the overall extraction method, it can more accurately grasp the dynamic changes of regional crops, which is of great significance for optimizing resource allocation and comprehensive agricultural development in the region. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to make the technical solutions of the embodiments of the present invention clearer, the drawings required for describing the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and ordinary technicians in this field can also obtain other drawings based on these drawings without creative work.

[0059] Figure 1 A flowchart of a planting structure recognition system and method based on GEE multiple extraction provided by the present invention;

[0060] Figure 2 The basic element extraction map of land use type provided by the present invention;

[0061] Figure 3 The NDVI spectrum curve graph of land use type provided by the present invention;

[0062] Figure 4 NDVI spectrum curve graph of the planting structure provided by the present invention;

[0063] Figure 5 A comparison diagram of the land use type extraction details provided by the present invention;

[0064] Figure 6 A spatial distribution diagram of the planting structure provided by the present invention; DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are 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.

[0066] Example 1 Reference Figure 1 As shown, a planting structure recognition system and method based on GEE multiple extraction includes the following steps:

[0067] Step 1: Extract basic elements of land use types using high-resolution images;

[0068] Obtain the latest high-resolution image (Tiantu) data and import it into ArcGIS layer as base map;

[0069] Villages and water bodies were selected as the basic elements of land use types. Village and water body features were created in ArcGIS and extracted through visual interpretation. The results are as follows: Figure 2 .

[0070] Step 2: Calculate the NDVI values ​​for different land use types and different crops;

[0071] The average NDVI values ​​of the sample points of different land use types from April to September were extracted to draw spectral curves to analyze their characteristics and differences. The results are as follows: Figure 3 ;

[0072] The average NDVI values ​​of the sample points with different planting structures from April to September were extracted to draw spectral curves to analyze their characteristics and differences. The results are as follows: Figure 4 .

[0073] Step 3: Call Sentinel-2 remote sensing images in GEE and import sample points and boundaries;

[0074] Select L2A data of Sentinel-2 that has undergone orthorectification, geometric correction, and atmospheric correction, and directly call it from the GEE cloud platform to select images with cloud cover less than 10%;

[0075] Field surveys were conducted to set up sample points for each land use type, including ditches, wasteland, and cultivated land. The cultivated land sample points were then refined to obtain separate sample points for corn, sunflower, and other crops. The principle of sample point selection was to be as uniform and representative as possible. Field surveys were conducted every month from April to September, and the field coordinates were recorded using GPS technology.

[0076] The vector files of the study area boundary and sample points were imported into GEE, and the image data were cropped according to the shape of the study area.

[0077] Step 4: Extract the reflectance and NDVI value of the sample point band as characteristic parameters;

[0078] Since the B1 and B9 bands are coastal aerosol and water vapor bands, respectively, they have no practical significance, so the B2 (blue), B3 (green), B4 (red), B5 (red edge 1), B6 ​​(red edge 2), B7 (red edge 3), B8 (near infrared), B8A (narrow near infrared), B11 (shortwave infrared 1) and B12 (shortwave infrared 2) bands are selected for band reflectance extraction;

[0079] The extracted band reflectance and NDVI values ​​are used as characteristic parameters.

[0080] Step 5: The machine learning model extracts the land use type after removing the basic elements;

[0081] Since May to September is the main growth period of crops, during this period, the cultivated land is covered with crops, and the remote sensing images are obviously different from wasteland and road ditches, so the minimum value images that can better display crops from May to September are synthesized to extract land use types;

[0082] GEE combined with gradient boosting tree, random forest, decision tree, support vector machine and naive Bayes machine learning models with feature parameters were used to extract land use types after removing the basic elements of land use types.

[0083] Step 6: Compare the accuracy of different models to obtain the optimal land use type extraction model;

[0084] The overall accuracy (OA), producer accuracy (PA), user accuracy (UA) and Kappa coefficient were selected to evaluate the extraction accuracy of the five machine learning models. It was found that the gradient boosting tree model had higher accuracy and the best classification effect, as shown in Table 1.

[0085] A detailed comparison of land use types extracted using the gradient boosting tree model is shown in the following figure. Figure 5 .

[0086] Table 1 Accuracy evaluation of land use type classification model

[0087] Step 7: Extract land use type of cultivated land attributes;

[0088] The extracted land use type raster data were imported into ArcGIS, and cultivated land was extracted by attributes and converted into vector data.

[0089] Step 8: The machine learning model extracts the planting structure of the cultivated land;

[0090] The extracted cultivated land vector data were imported into GEE. Different crops showed different NDVI in July and August, so this period was selected as the extraction time and the corresponding median image was synthesized.

[0091] The gradient boosting tree, random forest, decision tree, support vector machine and naive Bayes machine learning model with added feature parameters were used to extract the planting structure of cultivated land data.

[0092] Step 9: Compare the accuracy of different models, obtain the optimal implant structure extraction model and identify it;

[0093] The overall accuracy (OA), producer accuracy (PA), user accuracy (UA) and Kappa coefficient were selected to evaluate the extraction accuracy of the five machine learning models, and it was found that the gradient boosting tree model had a higher accuracy in identifying planting structures.

[0094] The identified data were imported into ArcGIS for mapping to obtain the spatial distribution map of the planting structure. Figure 6 .

[0095] The present invention proposes a planting structure recognition system and method based on GEE multiple extraction. The basic elements of land use types are extracted using high-resolution images; the gradient boosting tree, random forest, decision tree, support vector machine and naive Bayes machine learning models are selected to extract the land use types after removing the basic elements; the extraction accuracy of the five machine learning models is evaluated to obtain the model with the best land use type classification effect; the land use type of cultivated land attributes is extracted, and the accuracy of the five machine learning models is evaluated to obtain the model with the best planting structure classification effect, thereby achieving fine recognition of planting structures. The problem of being unable to accurately identify planting structures due to the extremely similar characteristics of some different land objects is solved.

[0096] The principles and implementation methods of the present invention are described in this article using specific examples. Each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. For any technician familiar with the technical field within the technical scope disclosed by the present invention, equivalent replacement or change according to the technical solution and inventive concept of the present invention should be covered within the protection scope of the present invention.

Claims

1. A planting structure recognition system and method based on GEE multiple extraction, characterized in that: The following steps are involved: S1. Extract basic elements of land use types using high-resolution images; S2. Calculate the NDVI values ​​for different land use types and different planting structures; S3. Call Sentinel-2 remote sensing images in GEE and import sample points and boundaries; S4, extracting the band reflectance and NDVI value of the sample point as characteristic parameters; S5, machine learning model extracts land use types after removing basic elements; S6. Compare the accuracy of different models to obtain the optimal land use type extraction model; S7, extracting land use type of cultivated land attributes; S8, machine learning model extracts the planting structure of cultivated land; S9. Compare the accuracy of different models to obtain the optimal implant structure extraction model and identify it.

2. A planting structure recognition system and method based on GEE multiple extraction according to claim 1, characterized in that: In S1, the specific steps of extracting the basic elements of land use types are as follows: S1.

1. Obtain the latest Tiandi Map image data as the base map; S1.

2. Select the basic elements of land use types, create features in ArcGIS and extract them through visual interpretation.

3. A planting structure recognition system and method based on GEE multiple extraction according to claim 1, characterized in that: In S2, the NDVI calculation formula is: Where NIR is the reflectivity of the B8 near-infrared band; Red is the reflectivity of the B4 red light band.

4. A planting structure recognition system and method based on GEE multiple extraction according to claim 1, characterized in that: In S3, the specific steps for calling remote sensing images are as follows: S3.

1. Image call: Select L2A-level data that has been orthorectified, geometrically corrected, and atmospherically corrected by Sentinel-2, and call it directly from the GEE cloud platform; S3.

2. Image selection rules: Screen images with cloud cover less than 10% and crop and synthesize the image data according to the shape of the study area.

5. The planting structure recognition system and method based on GEE multiple extraction according to claim 1 is characterized in that: In S5, the main principles of the machine learning model are as follows: S5.

1. Decision tree is a machine learning algorithm that makes decisions based on a tree structure. It performs a series of feature tests on the data and divides the data into different branches according to the test results, eventually forming a tree structure. Each internal node represents a test of a feature attribute, the branch represents the test result, and the leaf node represents the final classification result. When constructing a decision tree, information gain, information gain rate, Gini index and other indicators are usually used to select the optimal features for division. The indicator principle formula is as follows: Where D is the data set; Ent(D) represents the information entropy of the data set D; P k represents the proportion of category k in the data set D, |y| is the number of categories; Gain(D, a) is the information gain; Gain_ratio(D, a) is the information gain rate; Gini(D) is the Gini value; Gini_index(D, a) is the Gini index; a is a certain feature; V is the number of values ​​of feature a; D v is the data subset when feature a takes the value v; S5.

2. Random forest is a machine learning algorithm that consists of multiple decision trees. When constructing a random forest, multiple sample sets are first randomly extracted from the original training data. Each sample set is used to train a decision tree. When constructing each decision tree, a part of the features is also randomly selected for division. Finally, the final classification result is determined by voting. S5.

3. The gradient boosting tree model iteratively trains a series of decision trees and combines them to obtain the final prediction result. In each iteration, the algorithm builds a new decision tree based on the prediction error of the current model, so that the new decision tree can reduce the error of the current model. The main principle formula is as follows: F m+1 (x)=F m (x)+h m (x) In the formula, x is the eigenvector; x i is the feature vector of the i-th sample point; F m (x) is the model after m iterations; y i is the true value; L(y i , F m (x i )) is the loss function; r mi is the negative gradient of the loss function, and the training target of the new decision tree is determined by calculating the negative gradient; h m (x) is the newly constructed decision tree; S5.

4. Support vector machine is a binary classification model. Its basic idea is to find a hyperplane that can separate data points of different categories as correctly as possible and maximize the distance from the point closest to the hyperplane to the hyperplane. For the linearly separable case, the classification decision function is: f(x)=sign(w T x+b) In the formula, w is the normal vector; b is the intercept; x is the data point; For the case of linear inseparability, the kernel function is introduced to map the data to a high-dimensional space, so that the data becomes linearly separable in the high-dimensional space. The classification decision function is: In the formula, b is the intercept; a i is the Lagrange multiplier; y i is the category label; n is the number of samples; K(x i ,x) is the kernel function; S5.

5. Naive Bayes is a classification method based on Bayes’ theorem and the conditional independence assumption of features. For given input data, Naive Bayes calculates the probability of each feature appearing in each category, and then calculates the posterior probability that the data belongs to each category according to Bayes’ theorem. Finally, the data is assigned to the category with the largest posterior probability. The main principle formula is as follows: Where P(C|X) is the posterior probability of belonging to category C given feature X; P(X|C) is the probability of feature X appearing under category C; P(C) is the prior probability of category C; and P(X) is the marginal probability of feature X.

6. A planting structure recognition system and method based on GEE multiple extraction according to claim 1, characterized in that: In S6, the overall accuracy (OA), producer accuracy (PA), user accuracy (UA) and Kappa coefficient are used to compare the accuracy evaluation parameters of different machine learning models. The specific calculation formula is: In the formula, x is the number of validation samples; and is the total number of samples of the i-th category in the classification results and the total number of samples of the i-th category in the validation samples; X ii is the number in the i-th row and i-th column of the error matrix; it represents the number of correctly classified samples in the i-th category; k is the number of classification categories.

Images