Cultivated land extraction method based on remote sensing cloud computing platform and integrated machine learning technology

By adopting remote sensing cloud computing platform and integrated machine learning technology in arable land extraction technology, comprehensive feature sets are built and integrated learning is carried out, data processing and robustness problems in large-scale long-term land extraction are solved, and high-precision and Stable arable land extraction results are suitable for remote sensing data in different regions and time phases.

CN118918492APending Publication Date: 2024-11-08INST OF GEOGRAPHICAL SCI & NATURAL RESOURCE RES CAS
View PDF 0 Cites 8 Cited by

Patent Information

Application Number
CN202411095522.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-12
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing arable land extraction technology has problems such as difficulty in processing massive remote sensing data, insufficient feature extraction, insufficient algorithm robustness and inconsistent post-processing technology in large-scale and long-term land extraction, resulting in low classification accuracy and poor adaptability.

Method used

Using a method based on remote sensing cloud computing platform and integrated machine learning technology, a comprehensive feature set containing spectral index, topographic factors and texture features is constructed through pre-processing of multi-time phase remote sensing images and digital elevation model data, and combined with random forests and support vectors The computer algorithm performs integrated learning, performs weighted voting method integrated model results, and finally performs morphological post-processing to remove noise and fine pattern spots.

Benefits of technology

Improve classification accuracy and robustness, enhance the stability and adaptability of classification results, and can more accurately identify and extract cultivated land. It is suitable for remote sensing data in different regions and time phases, providing high-precision cultivated land management and monitoring support. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918492A_ABST
    Figure CN118918492A_ABST
Patent Text Reader

Abstract

The invention relates to a remote sensing surveying and mapping technology, in particular to a cultivated land extraction method based on a remote sensing cloud computing platform and an integrated machine learning technology, which comprises the following steps: S1, data preprocessing; s2, constructing a feature set; s3, selecting training sample points; s4, integrating a machine learning model; step S5, performing morphological post-treatment; and step S6, result precision verification. According to the method provided by the invention, the classification precision is improved, the stability and robustness of a classification result are enhanced, and feature set construction and model parameter setting are flexibly adjusted, so that the method can adapt to different geographical environments and seasonal changes, and accurate extraction of cultivated lands in different regions is realized; the method can be applied to remote sensing data of different areas and different time phases, and is suitable for large-scale long-time-sequence remote sensing cultivated land extraction and monitoring.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of remote sensing mapping technology, in particular to remote sensing mapping technology of cultivated land plots, and specifically to a cultivated land extraction method based on a remote sensing cloud computing platform and integrated machine learning technology. Background Art

[0002] The rapid and accurate extraction of cultivated land is an important basis for supporting the protection, development and utilization of cultivated land resources. Traditional cultivated land survey methods mainly rely on field surveys and manual mapping. Although this method has high accuracy, it has problems such as high cost, long time consumption and high labor intensity. In addition, field surveys are limited by terrain conditions and weather, and it is difficult to complete large-scale regional surveys in a short period of time. Earth exploration and remote sensing technology provides a reliable and cost-effective method for achieving long-term and large-scale cultivated land extraction. Conventional cultivated land extraction methods based on remote sensing images usually use remote sensing images for visual interpretation or simple threshold segmentation. These methods are highly dependent on the extraction target and require manual setting of parameters. They are suitable for areas with obvious land features. In addition, these methods have low accuracy and lack automated processing capabilities.

[0003] In order to improve the automation and accuracy of cultivated land extraction, machine learning technology has been introduced into remote sensing image classification. Cultivated land extraction has gradually transitioned from traditional local visual interpretation and pixel-based classification algorithms to a new stage of intelligent cultivated land extraction represented by machine learning and deep learning. Commonly used machine learning algorithms include decision trees (DT), support vector machines (SVM), and random forests (RF). These methods automatically classify ground object pixels in remote sensing images by constructing classifiers. Machine learning methods can process more complex multi-dimensional feature data and improve classification accuracy, but are susceptible to noise and changes in data distribution. Deep learning technology can automatically extract high-level deep features in images and reduce dependence on manually set features. Deep learning models such as convolutional neural networks (CNN) and U-Net networks can better capture complex features in remote sensing images and improve classification accuracy and robustness through hierarchical feature extraction. However, the training process of deep learning models is complex, the computing resource requirements are high, and the generalization ability of the model is poor, especially when migrated to other areas, the performance often decreases. In addition, due to the huge amount of data training and algorithm complexity, it is difficult to achieve large-scale and long-time series farmland extraction.

[0004] In general, farmland extraction technology has gradually developed from traditional field survey methods to automated processing using remote sensing technology, among which the following methods have been derived:

[0005] Contour-oriented farmland extraction method[1]

[0006] like Fig. 9As shown in the figure, a contour-oriented method based on the Richer Convolution Features (RCF) network is used to extract farmland contours from high-resolution remote sensing images. The method is divided into three stages, including data preparation, network training, and farmland extraction. In the data preparation stage, the high-resolution image data of the study area are annotated to form training, verification, and test data sets; in the network training stage, the RCF network is called, which is good at capturing scale-invariant edge features. Through multi-layer convolution and pooling operations, the RCF network can learn the farmland contour features in the image. During the training process, the network is optimized through the cross entropy loss function and weighted loss strategy; in the farmland extraction stage, the trained model is used to predict the farmland image to be extracted, generate the farmland plot contour, and generate the final farmland distribution result through manual correction.

[0007] Farmland extraction method based on machine learning[2]

[0008] like Fig.10 As shown in the figure, a machine learning algorithm was used to map the farmland and cultivated land range in the study area with high precision. The method used multispectral data (including eight bands such as SWIR 1, SWIR 2, near infrared, red light, green light, and blue light) and the obtained high-resolution training and test data, and analyzed them through pixel-level classification algorithms. Multi-temporal remote sensing data were preprocessed to generate cloud-free time series composite images; multi-file data cube (MFDC) was used to integrate image data and construct a 48-band time series data set; four machine learning algorithms, random forest, support vector machine (SVM), naive Bayes and classification and regression tree (CART), were selected for training and verification. The model parameters and samples were optimized through multiple iterations, and the optimized model was used to classify the entire study area to generate cultivated land range results.

[0009] Farmland extraction method based on deep learning[3]

[0010] like Fig.11 As shown in the figure, this method uses a deep learning model FracTAL ResUNet and introduces two uncertainty metrics to characterize the confidence in the process of cultivated land detection and boundary delineation. The FracTAL ResUNet network performs semantic segmentation on single-temporal images, predicts the existence, boundaries and distance to the nearest boundary of farmland, and averages the prediction results of multi-temporal images in time to integrate temporal clues and improve prediction accuracy. Through the hierarchical watershed segmentation method, the boundary mask is converted into a closed cultivated land boundary and the cultivated land extraction result is generated.

[0011] However, the above three methods still have the following disadvantages:

[0012] 1. For large-scale and long-time series ground feature extraction, it is difficult to achieve high-performance computing and processing of massive remote sensing data;

[0013] 2. The feature extraction is insufficient, and the constructed feature space cannot fully reflect the diversity and heterogeneity of the objects;

[0014] 3. The classification algorithm is not robust enough. Traditional machine learning methods are difficult to deal with the uncertainty in complex terrain situations, and the generalization ability of deep learning methods to other regions is not strong;

[0015] 4. The post-processing technology is not unified. There are often noise and misclassified patches in the classification results, which affect the final application effect. A unified post-processing technology flow has not yet been formed. Summary of the invention

[0016] In response to the problems raised by the background technology, the purpose of the present invention is to propose a cultivated land extraction method based on remote sensing cloud computing platform and integrated machine learning technology. This method can solve the problems of difficulty in processing remote sensing big data, insufficient feature extraction, insufficient algorithm robustness, and inconsistent post-processing technology in cultivated land extraction research.

[0017] To achieve the above object, the present invention adopts the following technical solutions:

[0018] A method for extracting cultivated land based on a remote sensing cloud computing platform and integrated machine learning technology includes the following steps:

[0019] Step S1, data preprocessing; data preprocessing; obtaining multi-temporal remote sensing images and digital elevation model data of the target area; taking the median of the band information of the multi-temporal remote sensing images grid by grid, synthesizing the multi-temporal remote sensing images into a comprehensive image without cloud coverage, mosaicking and cropping the synthesized comprehensive image to ensure that the comprehensive image covers the target area; filling and smoothing the digital elevation model data to remove outliers and noise in the data, obtaining smoothed elevation data, splicing and cropping the smoothed elevation data to ensure that the elevation data covers the target area;

[0020] Step S2, constructing a feature set: using the comprehensive image and elevation data processed in step S1, constructing a comprehensive feature set including spectral index, terrain factor and texture feature; wherein the texture feature is obtained by the following method: performing band calculation on the comprehensive image processed in step S1 to obtain a grayscale image, using a spatial grayscale co-occurrence matrix method to calculate specific texture feature parameters of the grayscale co-occurrence matrix, and selecting the specific texture feature parameters as the texture feature;

[0021] Step S3, selecting training sample points; visualizing the comprehensive image processed in step S1, identifying and marking cultivated land sample points and non-cultivated land sample points through manual visual interpretation;

[0022] Step S4, machine learning model integration; combining the comprehensive feature set obtained in S2 with the sample points obtained in S3 to form a high-dimensional matrix for input model training; using two different classification algorithms to train the models respectively, obtaining two different model results, integrating the results of the two models using a weighted voting method, and merging the results of the two models into an integrated learning model by weighted averaging, and the integrated learning model outputs the cultivated land extraction result;

[0023] Step S5, morphological post-processing: performing spatial filtering operation and area threshold screening on the cultivated land extraction results to remove small and broken patches in the cultivated land extraction results;

[0024] Step S6, result accuracy verification; use the confusion matrix to evaluate the cultivated land extraction classification results processed in step S5, and calculate the overall classification accuracy and Kappa coefficient; quantify the accuracy of the cultivated land extraction classification results by calculating the overall accuracy and Kappa coefficient.

[0025] In step S1, before the multi-temporal remote sensing images are synthesized into a comprehensive image, a cloud detection and removal step is required to detect the cloud cover in the multi-temporal remote sensing images using a cloud detection algorithm, and remove areas with high cloud cover.

[0026] In the step S2, the spectral index includes: normalized vegetation index, enhanced vegetation index, bare soil index, soil adjusted vegetation index and normalized water index; the above spectral index is obtained by the following method: spectral bands containing rich ground object information, namely blue light band, green light band and red light band are screened out from the comprehensive image processed in step S1, the screened spectral bands are linearly and nonlinearly combined, and the above spectral index is calculated by mathematical formula.

[0027] In step S2, terrain factors are calculated using the elevation data processed in step S1, and the terrain factors include elevation, slope and slope aspect.

[0028] In step S2, the specific feature parameters include: angular second moment (ASM), entropy (ENT), contrast (CON), inverse difference moment (IDM), correlation (CORR), dispersion (VAR), average value sum (SAVG), dispersion sum (SVAR), and entropy sum (SENT).

[0029] In step S4, the two different classification algorithms are random forest algorithm and support vector machine algorithm.

[0030] The step S5 is specifically as follows:

[0031] S5.1. Perform spatial filtering on the cultivated land extraction results. By spatially smoothing the image, the noise and isolated points in the cultivated land extraction results are removed, thereby improving the overall continuity and consistency of the cultivated land extraction results.

[0032] S5.2, further remove the fragmented patches, process the small patches in the cultivated land extraction results, remove the discontinuous small areas, and enhance the visual effect and practical application value of the cultivated land extraction results;

[0033] S5.3. Convert the raster cultivated land extraction results into vectors and remove vector patches with too small an area.

[0034] The present invention has the following beneficial effects:

[0035] 1. Improve classification accuracy

[0036] Through the preprocessing of multi-temporal remote sensing data and elevation data, clouds and noise were removed to ensure the clarity and accuracy of the data. In the feature set construction stage, the spectral index, terrain factors and texture features were combined to fully enrich the input features of the classification model, which can better reflect the actual situation of the objects, thereby improving the accuracy of the classification model. In the machine learning model stage, two algorithms, random forest and support vector machine, were used. The weighted voting method was used for ensemble learning, which effectively combined the advantages of the two algorithms and further improved the accuracy of the classification results. Finally, through morphological post-processing, the fine patches and noise were removed, making the classification results more continuous and consistent.

[0037] 2. Enhance the stability and robustness of classification results

[0038] The random forest model is resistant to noise and not prone to overfitting, and the support vector machine model performs well in high-dimensional space. The results of the two models are integrated by weighted voting, which not only improves the classification accuracy, but also enhances the stability and robustness of the classification results. Especially in the face of complex distribution of objects and diversified spectral characteristics, the classification method of the present invention shows strong adaptability, can effectively distinguish different categories of objects, and reduce classification errors.

[0039] 3. Improve farmland management and monitoring capabilities

[0040] The high-precision cultivated land classification results provide reliable data support for cultivated land management and monitoring. This technical solution can not only accurately identify and extract cultivated land, but also provide important reference for land use change monitoring, agricultural production management and environmental protection, and assist relevant departments to timely grasp the distribution and changes of cultivated land and make scientific decisions.

[0041] 4. Remote sensing data adapted to different regions and time phases

[0042] The method of the present invention has strong adaptability and versatility, can be applied to remote sensing data of different regions and different phases, and is suitable for large-scale and long-time series remote sensing cultivated land extraction and monitoring. By flexibly adjusting the feature set construction and model parameter settings, it can adapt to different geographical environments and seasonal changes, and realize accurate extraction of cultivated land in different regions. This feature makes the technical solution of the present invention have wide application potential worldwide, and can provide effective solutions for cultivated land monitoring and management in different regions. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of the present invention.

[0044] Figure 2 This is a remote sensing image diagram of an embodiment.

[0045] Figure 3 Graph showing the distribution of training samples in the embodiment.

[0046] Figure 4 It is a comparison chart of the cultivated land extraction results of a local plot in the embodiment.

[0047] Figure 5 It is an overall comparison diagram of the cultivated land extraction results of the embodiment.

[0048] Figure 6 This is a schematic diagram of the random forest model.

[0049] Figure 7 The schematic diagram of the support vector machine model.

[0050] Figure 8 Schematic diagram of the integrated learning model.

[0051] Fig. 9 Flowchart of contour-oriented farmland extraction method.

[0052] Fig.10 Flowchart of the cultivated land extraction method based on machine learning.

[0053] Fig.11 Flowchart of the farmland extraction method based on deep learning. DETAILED DESCRIPTION

[0054] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments.

[0055] The present invention proposes a method for extracting cultivated land based on a remote sensing cloud platform and integrated machine learning technology, which achieves high-precision classification and extraction of cultivated land through multi-step processing and analysis. The target areas selected in the embodiment are Selenge Province and Darkhan-Uul Province in Mongolia, where the water conditions are relatively good. The remote sensing images selected are the cloud-free images of Sentinel-2 from May 1 to October 1, 2023. Figure 2As shown. The northern region of Mongolia is rich in land resources. In addition to traditional animal husbandry, the development and utilization of arable land resources have been strengthened in recent years. The method provided by the present invention will be used to extract arable land in Selenge Province and Darkhan-Uul Province in 2023.

[0056] like Figure 1 As shown, the method for extracting cultivated land plots includes the following steps:

[0057] Step S1: Data preprocessing

[0058] The data preprocessing stage involves the processing of multi-temporal remote sensing data and digital elevation model (DEM) data of the target area. To avoid the influence of winter snow accumulation, the time range of multi-temporal remote sensing images of the target area is controlled between May and October. Through the cloud detection and removal steps, the cloud detection algorithm is used to detect the cloud cover in the multi-temporal remote sensing images, and the areas with high cloud cover are removed to obtain a clearer image data set. Then, the image is synthesized, and the median of the multi-temporal remote sensing image band information is taken grid by grid and synthesized to obtain a cloud-free comprehensive image to ensure the integrity and continuity of the comprehensive image. After that, the synthesized comprehensive image is mosaicked and cropped to ensure that the comprehensive image covers the target area. The digital elevation model data processing involves filling and smoothing operations to remove outliers and noise in the data and obtain smooth elevation data. Similar to the comprehensive image, the processed elevation data is spliced ​​and cropped to ensure that the elevation data covers the target area.

[0059] Step S2: Feature set construction

[0060] The feature set construction stage includes spectral band selection, spectral index calculation, terrain factor calculation and texture feature acquisition; the spectral bands of interest are screened out from the comprehensive image obtained in step S1, as shown in Table 1, namely Band2, Band3, and Band4, which contain rich ground feature information;

[0061] Table 1 Spectral band description

[0062]

[0063] Each spectral band is linearly and nonlinearly combined, and various spectral indices are calculated by mathematical formulas to mine deeper spectral information. As shown in Table 2, the spectral indices include the normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), bare soil index (BSI), soil adjusted vegetation index (SAVI) and normalized water index (NDWI). These indices can reflect the spectral characteristics of different types of land objects and increase the distinction between classes.

[0064] Table 2 Spectral index description

[0065]

[0066] In Table 2, NIR is the reflectivity of the near infrared band, RED is the reflectivity of the red band; SWIR is the reflectivity of the short-wave infrared band; B=BLUE is the reflectivity of the blue band; GREEN is the reflectivity of the green band;

[0067] The terrain factors, including elevation, slope and aspect, are calculated using digital elevation model data. Terrain factors play an important role in the classification of land features.

[0068] The grayscale image is obtained by band calculation on the composite image processed in step S1, and texture features are extracted from the grayscale composite image; specifically, the grayscale co-occurrence matrix method is used to calculate various features of the grayscale co-occurrence matrix; the spatial grayscale co-occurrence matrix (GLCM) is a method for extracting texture features based on statistics, which describes the situation in which the grayscale of adjacent pixels in a certain direction or two pixels at a certain distance in a local area or the entire area of ​​the image remains unchanged or changes, and then calculates the probability distribution of the grayscale value change, which reflects the texture features in different directions and distances in the remote sensing image;

[0069] From GLCM, 14 eigenvalues ​​such as angular second moment, entropy, contrast, correlation, and inverse difference moment can be extracted. In order to obtain information about the type of features through texture, grayscale distribution, spatial structure, etc., the present invention selects nine texture feature parameters including angular second moment (ASM), entropy (ENT), contrast (CON), inverse difference moment (IDM), correlation (CORR), dispersion (VAR), average value (SAVG), dispersion (SVAR), and entropy (SENT) to add to the classification feature set. The specific description is shown in Table 3.

[0070] Table 3 Description of GLCM texture features

[0071]

[0072] In Table 3, P(i,j) is the value of the gray level co-occurrence matrix at position (i,j), μ is the mean gray level of the image, and P s (i) is the distribution of gray levels and, S is the mean, and N is the number of gray levels in the image;

[0073] Through the above steps, a comprehensive feature set containing spectral, topographic and texture information was constructed, providing rich information for subsequent model training.

[0074] Step S3: Selection of training sample points

[0075] The training sample point selection stage is mainly completed by visualizing the multi-temporal remote sensing images and performing manual visual interpretation; specifically, the comprehensive image processed in step S1 is visualized, and the RGB true color synthetic visualization image is used to make the different ground features in the image intuitive and clear for easy identification. The cultivated land sample points and non-cultivated land sample points are identified and marked in the color synthetic image by manual visual interpretation, or the distribution positions of the sample points are mapped separately after manual identification (such as Figure 3 As shown in the figure, the cultivated land sample point label is "1", and the non-cultivated land sample point label is "0". The representative sample points are accurately selected to provide high-quality training data for the subsequent machine learning model training;

[0076] Step S4: Machine Learning Model Integration

[0077] Before training the machine learning model, it is necessary to construct a training data set, combine the feature set obtained by S2 with the sample points obtained by S3, and form a high-dimensional matrix for input into the model training;

[0078] In the machine learning model stage, the random forest algorithm and support vector machine algorithm are used to train the model respectively; the random forest model is trained by building multiple decision trees and combining the results of each decision tree for classification, which has high accuracy and stability; Figure 6 As shown, suppose there are T decision trees h1,h2,…,h T , for the input sample x, the classification result H(x) of the random forest is the mode of all decision tree classification results, that is:

[0079] H(x)=mode{h1(x),h2(x),…,h T (x)}

[0080] The support vector machine model separates data points of different categories by finding the best hyperplane in high-dimensional space. It is also a commonly used classification algorithm. Figure 7 As shown, given a training data set (x1, y1), (x2, y2), …, (x n ,y n ), where x i is the eigenvector, y i ∈{-1,1} is the category label, and SVM finds the optimal hyperplane by solving the following optimization problem, namely:

[0081]

[0082] subject to you i (w·x i +b)≥1-ξ i ,ξ i ≥0

[0083] Where w is the normal vector of the hyperplane, b is the bias, ξ is the slack variable, and C is the regularization parameter;

[0084] The results of the two models are integrated by weighted voting method. The results of the random forest model and the support vector machine model are weighted averaged to form an integrated learning model, which outputs the cultivated land extraction results, thereby improving the accuracy and robustness of classification. Figure 8 As shown, the classification result of random forest h rf (x) and the classification result of the support vector machine h svm (x), the prediction result of weight integration H weighted (x) is the rounded result of the weighted sum of the two:

[0085] H weighted (x) = round(w rf ·h rf (x)+w svm ·h svm (x)

[0086] Among them, w rf and w svm are the weights of random forest and support vector machine, satisfying w rf +w svm =1; in the present invention, w rf and w svm Set to 0.7 and 0.3 respectively;

[0087] The integrated learning model combines the advantages of the two models to obtain more accurate farmland extraction results and improve the overall accuracy and reliability of farmland extraction results.

[0088] Step S5: Morphological post-processing

[0089] The morphological post-processing stage mainly includes two steps: spatial filtering operation and removal of fine and fragmented patches. First, a spatial filtering operation is performed on the cultivated land extraction results. By spatially smoothing the image, the noise and isolated points in the cultivated land extraction results are removed, thereby improving the overall continuity and consistency of the cultivated land extraction results; further remove small and fragmented patches, use morphological operations such as dilation and erosion to process the fine patches in the cultivated land extraction results, remove discontinuous small areas, and enhance the visual effect and practical application value of the classification results; convert the raster cultivated land extraction results into vectors, and remove vector patches with too small an area; after morphological post-processing, the comparison between the original extraction results of the embodiment and the morphological post-processing results is as follows: Figure 4 , Figure 5As shown, it can be seen that these post-processing steps can effectively improve the quality of cultivated land extraction results, make the cultivated land extraction results more consistent with the spatial distribution characteristics of actual objects, and improve the practicality and accuracy of cultivated land extraction results;

[0090] Step S6: Result accuracy verification

[0091] In the accuracy verification stage, the confusion matrix is ​​used to evaluate the cultivated land extraction results, and the overall accuracy (OA) and Kappa coefficient of the classification are calculated; the confusion matrix can intuitively display the correctness and error of the classification results, and quantify the accuracy of the classification results by calculating the overall accuracy and Kappa coefficient; the final classification results are exported to tiff format for further analysis and application; through accuracy verification, the performance of the model can be evaluated, and the model can be adjusted and optimized according to the accuracy evaluation results to improve the reliability and accuracy of the model. The overall accuracy is calculated according to the following formula:

[0092]

[0093] Among them, C ii is the element in the ith row and ith column of the confusion matrix, indicating the number of correct classifications, k is the total number of categories, and N is the total number of samples, that is, the sum of all elements of the confusion matrix.

[0094] R i is the sum of the i-th row of the confusion matrix, indicating the total number of actual samples of the i-th class, C i is the sum of the i-th column of the confusion matrix, indicating the total number of predicted samples of the i-th class. Then, the Kappa coefficient is calculated as follows:

[0095]

[0096] Among them, P e is random consistency, R i is the sum of the i-th row of the confusion matrix, indicating the total number of actual samples of the i-th class, C i is the sum of the i-th column of the confusion matrix, indicating the total number of predicted samples of the i-th class.

[0097] The technical effects of the present invention include the following three aspects:

[0098] 1. Comprehensive feature construction technology

[0099] A comprehensive feature set construction method was developed, covering spectral features, terrain features and texture features, including blue, green, red, NDVI, EVI, BSI, SAVI, NDWI, elevation, slope, aspect and several texture features. The important spectral bands in the remote sensing images were screened and a variety of spectral indices were calculated to reflect the spectral characteristics of different objects such as vegetation, water bodies, and bare land. Terrain factors such as slope and aspect were calculated using elevation data to assist in identifying objects on different terrains. Texture features were extracted from the images, and the gray level co-occurrence matrix (GLCM) method was used to calculate multiple features such as contrast, entropy, and uniformity of the gray level co-occurrence matrix to further improve the accuracy of cultivated land extraction results.

[0100] 2. Machine Learning Ensemble Learning Technology

[0101] An ensemble learning method for machine learning models is proposed. Classification is performed using two algorithms, random forest and support vector machine, and the results are integrated using weighted voting. The random forest model improves the robustness and noise resistance of the model by constructing multiple decision trees and using diverse training data; the support vector machine model searches for the best hyperplane in high-dimensional space to achieve accurate classification. The ensemble learning method combines the advantages of the two machine learning models and integrates the classification results through weighted voting to further improve the classification accuracy and stability.

[0102] 3. Morphological post-processing technology

[0103] Morphological post-processing methods are introduced to perform spatial filtering, smoothing and vector-raster conversion on the cultivated land extraction results in order to remove noise and fine patches, improve the continuity and consistency of the classification results, ensure the high quality of the classification results, enhance the visual effect and practical application value of the image, and conform to the spatial distribution characteristics of actual objects.

[0104] Parts of the present invention not described in detail are prior art.

Claims

1. A method for extracting cultivated land based on remote sensing cloud computing platform and integrated machine learning technology, characterized by: The following steps are involved: Step S1, data preprocessing; obtaining multi-temporal remote sensing images and digital elevation model data of the target area; taking the median of the band information of the multi-temporal remote sensing images grid by grid, synthesizing the multi-temporal remote sensing images into a comprehensive image without cloud coverage, mosaicking and cropping the synthesized comprehensive image to ensure that the comprehensive image covers the target area; filling and smoothing the digital elevation model data to remove outliers and noise in the data to obtain smooth elevation data, splicing and cropping the smoothed elevation data to ensure that the elevation data covers the target area; Step S2, constructing a feature set: using the comprehensive image and elevation data processed in step S1, constructing a comprehensive feature set including spectral index, terrain factor and texture feature; wherein the texture feature is obtained by the following method: performing band calculation on the comprehensive image processed in step S1 to obtain a grayscale image, using a spatial grayscale co-occurrence matrix method to calculate specific texture feature parameters of the grayscale co-occurrence matrix, and selecting the specific texture feature parameters as the texture feature; Step S3, selecting training sample points; visualizing the comprehensive image processed in step S1, identifying and marking cultivated land sample points and non-cultivated land sample points through manual visual interpretation; Step S4, machine learning model integration; combining the comprehensive feature set obtained in S2 with the sample points obtained in S3 to form a high-dimensional matrix for input model training; using two different classification algorithms to train the models respectively, obtaining two different model results, integrating the results of the two models using a weighted voting method, and merging the results of the two models into an integrated learning model by weighted averaging, and the integrated learning model outputs the cultivated land extraction result; Step S5, morphological post-processing: performing spatial filtering operation and area threshold screening on the cultivated land extraction results to remove small and broken patches in the cultivated land extraction results; Step S6, result accuracy verification; using the confusion matrix to evaluate the cultivated land extraction results processed in step S5, and calculating the overall classification accuracy and Kappa coefficient; quantifying the accuracy of the cultivated land extraction results by calculating the overall accuracy and Kappa coefficient.

2. The method for extracting cultivated land plots according to claim 1, characterized in that: In step S1, before the multi-temporal remote sensing images are synthesized into a comprehensive image, a cloud detection and removal step is required to detect the cloud cover in the multi-temporal remote sensing images using a cloud detection algorithm, and remove areas with high cloud cover.

3. The method for extracting cultivated land plots according to claim 1, characterized in that: In the step S2, the spectral index includes: normalized vegetation index, enhanced vegetation index, bare soil index, soil adjusted vegetation index and normalized water index; the above spectral index is obtained by the following method: spectral bands containing rich ground object information, namely blue light band, green light band and red light band are screened out from the comprehensive image processed in step S1, the screened spectral bands are linearly and nonlinearly combined, and the above spectral index is calculated by mathematical formula.

4. The method for extracting cultivated land plots according to claim 1, characterized in that: In step S2, terrain factors are calculated using the elevation data processed in step S1, and the terrain factors include elevation, slope and slope aspect.

5. The method for extracting cultivated land plots according to claim 1, characterized in that: In step S2, the specific feature parameters include: angular second moment (ASM), entropy (ENT), contrast (CON), inverse difference moment (IDM), correlation (CORR), dispersion (VAR), average value sum (SAVG), dispersion sum (SVAR), and entropy sum (SENT).

6. The method for extracting cultivated land plots according to claim 1, characterized in that: In step S4, the two different classification algorithms are random forest algorithm and support vector machine algorithm.

7. The method for extracting cultivated land plots according to claim 1, characterized in that: The step S5 is specifically as follows: S5.

1. Perform spatial filtering on the cultivated land extraction results. By spatially smoothing the image, the noise and isolated points in the cultivated land extraction results are removed, thereby improving the overall continuity and consistency of the cultivated land extraction results. S5.2, further remove the fragmented patches, process the small patches in the cultivated land extraction results, remove the discontinuous small areas, and enhance the visual effect and practical application value of the cultivated land extraction results; S5.

3. Convert the raster cultivated land extraction results into vectors and remove vector patches with too small an area.

Citation Information

Cited By

  • Irrigation area ground feature element ensemble learning classification method and system

    CN120147873A

  • Remote sensing image classification post-processing spatial filtering method considering environment similarity

    CN120219967A

  • Complex farming area idle cultivated land extraction method and device, model training method and device, storage medium and terminal

    CN120236096A

  • Space-time deep learning-based fragmented cultivated land remote sensing extraction method

    CN120339845A

  • Self-adaptive multi-scene cultivated land parcel refining segmentation method and device based on remote sensing image

    CN120411496A