A method for extracting a plantation forest from multiple remote sensing sources

Abstract

The present application relates to a kind of artificial forest multi-source remote sensing extraction method, belong to the field of remote sensing data detection, comprising: obtaining multi-source remote sensing data and pre-processing, obtain pre-processing multi-source remote sensing data;Respectively obtain training sample and verification sample, according to verification sample screening training sample to obtain first sample set;According to first sample set, respectively train several classifiers to obtain several first classifiers, and first sample set is classified to obtain first classification result;Sample enhancement is carried out to minority classification result in first classification result to obtain second sample set;According to second sample set, respectively train several first classifiers, corresponding to obtain several second classifiers;Through several second classifiers, respectively classify pre-processing multi-source remote sensing data, and the final classification result is obtained by decision fusion to second classification result.The present application is based on multi-source remote sensing data, according to sample enhancement and decision fusion of classification result, improve the recognition ability to minority class sample.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing data detection, specifically relating to a method for multi-source remote sensing extraction of plantation forests. Background Technology

[0002] With the rapid development of remote sensing technology and geographic information systems, large-scale forest monitoring and classification using remote sensing data has become an important means of forestry management. Remote sensing technology, by acquiring large-area, high-frequency surface information, provides reliable data support for studying the spatial distribution, dynamic changes, and relationship with environmental factors of forest resources. In particular, with the widespread availability of high-resolution, multispectral satellite imagery, remote sensing technology has been widely applied in fields such as forest classification, forest health monitoring, and carbon storage assessment.

[0003] Time-series remote sensing data, which involves observing changes in land surface targets over time through continuous multi-temporal imagery, has become an important tool in vegetation monitoring research. By analyzing the temporal changes of the Normalized Difference Vegetation Index (NDVI), the growth cycle, seasonal variations, and impacts of anomalous events on vegetation can be effectively captured. Particularly in forest classification, time-series analysis can identify and distinguish the growth patterns and seasonal characteristics of different forest types, thereby improving classification accuracy. Modern remote sensing technology no longer relies solely on single-type image data but has gradually developed methods for fusing multi-source remote sensing data, such as optical imagery, radar data, and digital elevation models. These data types each have advantages in spatial, spectral, and temporal resolution; by fusing multi-source data, richer information about the land surface can be obtained. In forest classification, multi-source data fusion technology can fully utilize spectral, spatial structure, and topographic information to enhance the ability to identify forest types from different dimensions. Meanwhile, with the improvement of computing power and the advancement of machine learning algorithms, machine learning-based remote sensing image classification technology has become a research hotspot. Random forests, support vector machines (SVMs), and deep learning algorithms have demonstrated powerful capabilities in processing high-dimensional remote sensing data. Especially when dealing with complex terrain and diverse land cover, machine learning algorithms can effectively utilize the features of training samples to build robust classification models.

[0004] Existing research combines long-term series data with machine learning methods to improve classification accuracy when distinguishing similar land cover types. For example, time-series analysis based on Harmonic Analysis of Time Series (HANTS) can remove cloud cover effects through harmonic analysis, reconstruct remote sensing time-series data, and use time-series trend features to distinguish different types of forests. These studies demonstrate that the application of long-term series data can significantly improve the accuracy of plantation identification and reduce uncertainty in classification. However, due to the significant difference in sample size between plantations and natural forests, complex forest classification tasks are greatly affected by data noise, and the multidimensional features in multi-source remote sensing data cannot be fully utilized during processing, resulting in low overall accuracy and stability of the final classification results. Therefore, a multi-source remote sensing extraction method for plantations is needed to improve the accuracy and stability of classification results. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a multi-source remote sensing extraction method for planted forests. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] This invention provides a method for multi-source remote sensing extraction of plantations, comprising:

[0007] Acquire multi-source remote sensing data;

[0008] Preprocessing of multi-source remote sensing data yields preprocessed multi-source remote sensing data.

[0009] Training samples and validation samples are obtained separately, and training samples are selected based on the validation samples to obtain the first sample set;

[0010] Several classifiers are trained based on the first sample set, resulting in several first classifiers. The first sample set is then classified using these first classifiers, resulting in several first classification results.

[0011] A second sample set is obtained by performing sample augmentation on a minority of the first classification results;

[0012] Several first classifiers are trained based on the second sample set, resulting in several second classifiers. The preprocessed multi-source remote sensing data are then classified using these second classifiers, resulting in several second classification results. Finally, the second classification results are fused to obtain the final classification result.

[0013] In one embodiment of the present invention, acquiring multi-source remote sensing data includes: acquiring Sentinel-2 time series images, Sentinel-1 time series images and Landsat time series images respectively.

[0014] In one embodiment of the present invention, preprocessing of multi-source remote sensing data to obtain preprocessed multi-source remote sensing data includes:

[0015] Cloud removal was performed on Sentinel-2 time series images, Sentinel-1 time series images, and Landsat time series images to obtain multispectral band information.

[0016] NDVI processing was performed on the cloud-removed Landsat time-series images to obtain the normalized vegetation index.

[0017] Harmonic regression was used to fit the normalized vegetation index (NWRI), and the statistical characteristics of the fitted NWRI were extracted. The statistical characteristics include: mean, variance, standard deviation, coefficient of variation, and corrected variance.

[0018] Multi-source remote sensing data also includes: topographic feature information and radar image data; multi-source feature integration is performed on multispectral band information, statistical features, topographic feature information and radar image data to obtain the integrated result;

[0019] Extract the integrated results to obtain texture features.

[0020] In one embodiment of the present invention, the formula for NDVI processing is:

[0021]

[0022] Wherein, NDVI is the Normalized Difference Vegetation Index; NIR is the reflectance value in the near-infrared band; and Red is the reflectance value in the red band.

[0023] In one embodiment of the present invention, the harmonic regression fitting model is as follows:

[0024]

[0025] in, The set of normalized vegetation indices after fitting and reconstructing; n is the number of harmonics; t j p is the time step, j = 1, 2, ..., N, where N is the number of observations in the time series; i and q i All are coefficients of trigonometric function terms in the Fourier series, f i x is the corresponding frequency; x0 is a constant term.

[0026] In one embodiment of the present invention, training samples and validation samples are obtained respectively, and training samples are filtered according to the validation samples to obtain a first sample set, including:

[0027] Training samples and validation samples are obtained separately, with the validation samples being forest layer data;

[0028] By comparing the training samples and the forest layer data, the training samples located within the forest layer data are selected as retained samples.

[0029] The retained samples are re-projected with labels to obtain the first sample set.

[0030] In one embodiment of the present invention, several classifiers are trained based on a first sample set to obtain several first classifiers, and the first sample set is classified by the several first classifiers to obtain a first classification result, including:

[0031] Minority class samples and majority class samples are selected from the first sample set as training sample sets, respectively.

[0032] Several classifiers are trained based on the training sample set; among them, the several classifiers include: random forest classifier, decision tree classifier and support vector machine classifier;

[0033] Adjust the proportion of minority class samples in the training sample set, and retrain several classifiers based on the adjusted training sample set until the evaluation metric is met, thus obtaining several first classifiers; where the evaluation metric is used to evaluate the classification accuracy of the several classifiers after training.

[0034] The first sample set is classified according to several first classifiers, and the corresponding first classification results are obtained.

[0035] In one embodiment of the present invention, the evaluation metrics include: confusion matrix results, overall precision, and recall.

[0036] The confusion matrix results include: true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN), which are used to determine the classification accuracy of several classifiers.

[0037] The formula for overall accuracy is:

[0038]

[0039] Where Overall Accuracy refers to the overall accuracy.

[0040] The formula for recall rate is:

[0041]

[0042] Here, Recall is the recall rate.

[0043] In one embodiment of the present invention, sample augmentation is performed on a minority of the several first classification results to obtain a second sample set, including:

[0044] The minority classification results from several primary classification results are compared with the forest layer data to determine the overlapping areas. The overlapping areas are as follows:

[0045] A overlap =A s ∩A f ;

[0046] Among them, A overlap A represents a small number of overlapping areas between classification results and forest layer data; s A region with a small number of classification results; f This refers to the region of the forest layer data;

[0047] A buffer is created centered on the few classification results located within the overlapping region. New sample points are generated within the buffer to achieve sample augmentation, and a second sample set is obtained based on the augmented samples.

[0048] The formula for the buffer is:

[0049] B i ={(x,y)|(xx) i ) 2 +(yy i ) 2 ≤r 2};

[0050] Where (x,y) is any point within the buffer; p i For the minority classification result, its coordinates are (x i ,y i ); B i To classify the minority results p i A buffer zone with center r and radius r;

[0051] The formula for the new sample point is:

[0052]

[0053] Where, p new For new sample points generated within the buffer; (x j ,y j ) represents the coordinates of the new sample point generated within the buffer; n new This represents the number of new sample points generated within the buffer.

[0054] In one embodiment of the present invention, several first classifiers are trained based on a second sample set, corresponding to several second classifiers, including:

[0055] Several first classifiers are trained based on the second sample set, and several second classifiers are obtained accordingly.

[0056] The preprocessed multi-source remote sensing data is classified using several secondary classifiers, resulting in several secondary classification results.

[0057] The decision fusion of several secondary classification results yields the final classification result, which includes:

[0058]

[0059] in, The final classification result is represented by z; z represents the number of pixels; mode is a function that calculates the category that appears most frequently among the three classification results. and The results are the classification results of several second classifiers, in sequence.

[0060] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0061] The multi-source remote sensing extraction method for planted forests of the present invention is based on multi-source remote sensing data. It performs sample enhancement based on the first classification result of the trained first classifier and generates new sample points in the buffer of minority class sample points. This not only effectively increases the number of minority class samples, but also ensures the reasonable distribution of new sample points, improves the classifier's ability to identify minority class samples, and solves the problem of class imbalance.

[0062] This invention utilizes the temporal features of NDVI processing and extracts trend and seasonal variation features from time series through harmonic regression fitting. This not only enhances the classifier's ability to capture vegetation changes but also improves the classifier's sensitivity to long-term and periodic changes, providing support for obtaining more accurate classification results.

[0063] This invention trains three different classifiers—random forest, decision tree, and support vector machine—and performs decision fusion to generate the final classification result. This fusion strategy can effectively integrate the advantages of different classifiers, overcome the bias problem that is prone to occur with a single classifier, and improve the overall accuracy, stability, and robustness of the final classification result.

[0064] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0065] Figure 1 This is a flowchart of a method for multi-source remote sensing extraction of artificial forests provided in an embodiment of the present invention;

[0066] Figure 2 This is a flowchart illustrating the process of a multi-source remote sensing extraction method for artificial forests provided in an embodiment of the present invention.

[0067] Figure 3 This is a flowchart of the sample enhancement method provided in an embodiment of the present invention;

[0068] Figure 4 This is a flowchart of the decision fusion method provided in the embodiments of the present invention. Detailed Implementation

[0069] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following describes in detail a method for multi-source remote sensing extraction of artificial forests based on the present invention, in conjunction with the accompanying drawings and specific embodiments.

[0070] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.

[0071] Example 1

[0072] like Figure 1 and Figure 2 As shown in the figure, this embodiment provides a method for multi-source remote sensing extraction of plantations, including:

[0073] S1: Acquire multi-source remote sensing data;

[0074] In an optional implementation, S1 includes: acquiring Sentinel-2 time-series images, Sentinel-1 time-series images, Landsat time-series images, multispectral remote sensing bands, terrain features, and radar image data, respectively.

[0075] S2: Preprocess the multi-source remote sensing data to obtain preprocessed multi-source remote sensing data;

[0076] In an optional implementation, S2 includes:

[0077] S2.1: Perform cloud removal processing on Sentinel-2 time series images, Sentinel-1 time series images and Landsat time series images respectively to obtain multispectral band information;

[0078] For example, cloud removal processing of Sentinel-2 time-series images includes: selecting Sentinel-2 images of 12 bands (B1-B12) from the COPERNICUS / S2_SR dataset on the GEE (Google Earth Engine) platform; using the COPERNICUS / S2_CLOUD_PROBABILITY dataset, masking the cloud-covered portions of the selected 12 bands of Sentinel-2 images according to a set cloud probability threshold (65%), and retaining pixels with cloud coverage below the threshold; applying masks of 20-meter and 60-meter bands to process edge data to address potential edge noise issues in some 10-meter resolution bands, thus ensuring image quality; and setting the time range of the images (January 1, 2018 to December 31, 2020) and the geometric extent of the study area to filter the images. The band parameters of the Sentinel-2 time-series images are shown in Table 1.

[0079] Table 1 Band parameters of Sentinel-2 time series images

[0080]

[0081] It is worth noting that the multi-source remote sensing extraction method for plantations in this embodiment achieves large-scale remote sensing data processing and analysis based on the GEE platform. The powerful computing and data storage capabilities of the GEE platform enable this multi-source remote sensing extraction method for plantations to perform efficient classification and analysis on large-scale, long-term data series. Compared to traditional local processing methods, this significantly improves data processing efficiency and reduces hardware costs and time consumption.

[0082] For example, cloud removal processing of Sentinel-1 time series images includes: selecting Sentinel-1 time series images containing VV and VH polarization information from the COPERNICUS / S1_GRD dataset; masking the noisy parts of the images using a set threshold to exclude anomalous reflectance values; and performing median synthesis processing on the images within the time range to reduce outlier interference in the time series and ensure the uniformity of the images.

[0083] For example, cloud removal processing of Landsat time-series imagery includes: selecting Landsat4, 5, 7, and 8 images from the Landsat image series and filtering them according to the region of interest and time range; applying corresponding scaling factors to the optical and thermal infrared bands to ensure the consistency of physical quantities in the Landsat images; linearly scaling and offset correction of the band values ​​according to the specific needs of each Landsat image series to obtain the actual surface reflectance and temperature values; generating a cloud mask using the quality assessment band (QA_PIXEL) in the imagery and excluding pixels covered by clouds, cloud shadows, and snow, retaining valid data without cloud cover. This processing ensures that the pixels in the imagery are free of interference, thereby improving the quality of image analysis.

[0084] S2.2: Perform NDVI processing on the Landsat time series images after cloud removal to obtain the normalized vegetation index;

[0085] In one optional implementation, the cloud-removed Landsat time-series imagery is processed with NDVI, and the Normalized Difference Vegetation Index (NDVI) is obtained by selecting the near-infrared and red bands. Subsequently, a time variable (in years, starting from 1970) is added to the imagery for subsequent time-series processing.

[0086] Specifically, the formula for NDVI processing is:

[0087]

[0088] Wherein, NDVI is the Normalized Difference Vegetation Index; NIR is the reflectance value in the near-infrared band; and Red is the reflectance value in the red band.

[0089] S2.3: Perform harmonic regression fitting on the normalized vegetation index and extract the statistical characteristics of the fitted normalized vegetation index. The statistical characteristics include: mean, variance, standard deviation, coefficient of variation and corrected variance.

[0090] In an optional implementation, the time series of the normalized vegetation index is reconstructed using a model fitted by harmonic regression. The harmonic regression model is as follows:

[0091]

[0092] in, The set of normalized vegetation indices after fitting and reconstructing; n is the number of harmonics; t j p is the time step, j = 1, 2, ..., N, where N is the number of observations in the time series; i and q i All are coefficients of trigonometric function terms in the Fourier series, fi x is the corresponding frequency; x0 is a constant term.

[0093] In an optional implementation, statistical features of the fitted normalized vegetation index are extracted. These statistical features are used to provide richer feature information for subsequent classifiers to obtain more accurate classification results.

[0094] For example, fitting the reconstructed time series of the normalized vegetation index. The mean can be expressed as the mean at all time points. The average value is calculated using the following formula:

[0095]

[0096] Where MeanNDVI is the mean, denoted by μ; and num is the total number of observations in the time series.

[0097] Variance is used to measure the time series of the reconstructed normalized vegetation index. The formula for extracting variance relative to the volatility of its mean MeanNDVI is:

[0098]

[0099] Where variance is the variance, expressed as σ. 2 express.

[0100] Standard deviation is the square root of variance and is used to describe the magnitude of fluctuation. The formula for extracting standard deviation is:

[0101]

[0102] Wherein, Standard Deviation is the standard deviation, denoted by σ.

[0103] The coefficient of variation (CV) is used to describe the time series of the reconstructed normalized vegetation index. The formula for extracting the coefficient of variation relative to its mean MeanNDVI is as follows:

[0104]

[0105] Where CV represents the standard deviation.

[0106] In a time series, the maximum and minimum values ​​represent the peak and lowest points of vegetation cover, respectively. The formulas for extracting the maximum and minimum values ​​are as follows:

[0107]

[0108] Where MaxNDVI is the maximum value, used MinNDVI represents the minimum value, expressed as: express.

[0109] Alternatively, other time-series feature extraction methods can be used to extract statistical features, such as wavelet transform, Fourier transform, and dynamic time warping.

[0110] S2.4: Multi-source remote sensing data also includes: topographic feature information and radar image data information; multi-source feature integration is performed on multispectral band information, statistical features, topographic feature information and radar image data information to obtain the integration result;

[0111] S2.5: Extract the integrated results to obtain texture features.

[0112] In one optional implementation, multi-source feature integration is performed on Sentinel-2 time-series images, Sentinel-1 time-series images, Landsat time-series images, multispectral remote sensing bands, topographic features, and radar image data from multi-source remote sensing data, and texture features are extracted based on the integration results.

[0113] For example, texture features are important characteristics of an image, representing the spatial variation of image gray levels. Texture features are extracted using the Gray-Level Co-occurrence Matrix (GLCM) to represent the joint probability distribution of gray levels between two pixels with a specific spatial relationship. Based on the GEE platform, the GLCM Texture function is used, which calculates the GLCM texture index around each pixel in each band, generating 18 texture feature bands for each texture feature, as shown in Table 2.

[0114] Table 2 Texture features and their meanings

[0115]

[0116]

[0117] Furthermore, the steps for extracting texture features using GLCM include:

[0118] Multi-source remote sensing data is converted into a grayscale image with U gray levels, ranging from 0 to U-1.

[0119] Construct a gray-level co-occurrence matrix G(x,y), where U×U is the size of the gray-level co-occurrence matrix G(x,y); (x,y) is a pixel pair consisting of two gray-level combinations; x is the gray level of the first pixel in the pixel pair; y is the gray level of the second pixel in the pixel pair; f is the distance between pixel pairs; and α is the direction of the pixel pair.

[0120] The gray-level co-occurrence matrix G(x,y) is normalized to obtain the normalized gray-level co-occurrence matrix G'(x,y).

[0121]

[0122] Where ∑∑G(x,y) is the sum of all elements in the normalized gray-level co-occurrence matrix G'(x,y);

[0123] Texture features are obtained based on the normalized gray-level co-occurrence matrix G'(x,y). These features include contrast, entropy, correlation, and average, with the following formulas:

[0124]

[0125]

[0126] Where CONTRAST is contrast; ENT is entropy; CORR is correlation; SAVG is sum-mean; μ i μ is the mean of the row indices. j σ is the mean of the column indexes; i σ is the standard deviation of the row index; j U is the standard deviation of the column index; s(k) It is the sum of probabilities of pixels in the gray-level co-occurrence matrix G(x,y) that satisfy x+y=k.

[0127] S3: Obtain training samples and validation samples respectively, and filter training samples based on validation samples to obtain the first sample set;

[0128] In an optional implementation, S3 includes:

[0129] S3.1: Obtain training samples and validation samples respectively. The validation samples are forest layer data.

[0130] For example, the training samples came from an online network system (https: / / survey.chinavegetation.cn / survey), which included vegetation-related training samples, and each sample recorded the attributes of the vegetation group or subgroup and its geographic coordinates. To ensure the accuracy of the samples, GlobeLan d30 data was used as validation samples. Developed by the National Geographic Information Center of China, this data covers land cover in 2020 and includes 10 major land cover types, such as forests, grasslands, wetlands, water bodies, and artificial surfaces. It provides reliable forest boundary information to verify the correctness of the training samples.

[0131] S3.2: Compare the training samples and the forest layer data, and select the training samples located within the forest layer data as retained samples;

[0132] For example, plantation and natural forest samples in the training dataset are compared with the forest layer of the GlobeLand30 data. Through filtering, training samples located within the GlobeLand30 forest layer are retained, while samples not within the forest layer are considered inaccurate (mistaken) and are discarded.

[0133] S3.3: Reproject the labels onto the retained samples to obtain the first sample set.

[0134] For example, after sample screening is completed, the retained samples undergo label reprojection. Specifically, the labels for natural forest samples are set to 0, and the labels for plantation forest samples are set to 1.

[0135] S4: Train several classifiers based on the first sample set to obtain several first classifiers, and classify the first sample set using the several first classifiers to obtain several first classification results;

[0136] In an optional implementation, S4 includes:

[0137] S4.1: Select minority class samples and majority class samples from the first sample set as training sample sets respectively;

[0138] Here, let K be the ratio of minority class to majority class samples. Class imbalance is considered to exist in the sample set when 0 < K ≤ 2 / 3. 70% of the minority class samples are selected as the training set, and 30% as the validation set. The proportion of majority class samples used as training samples varies, generally selected to achieve the highest recall. Correspondingly, after determining the number of majority class training samples, the number of majority class validation samples is determined as follows: majority class training samples : majority class validation samples = 7:3.

[0139] S4.2: Train several classifiers based on the training sample set. The classifiers include: random forest classifier, decision tree classifier and support vector machine classifier.

[0140] S4.3: Adjust the proportion of minority class samples in the training sample set, and retrain several classifiers based on the adjusted training sample set until the evaluation index is met, thus obtaining several first classifiers; where the evaluation index is used to evaluate the classification accuracy of the trained classifiers.

[0141] S4.4: Classify the first sample set according to several first classifiers to obtain the corresponding first classification results.

[0142] For example, to address the class imbalance problem, a smaller number of samples (minority class samples), such as plantation forest samples, are selected from the first sample set as a baseline, and these samples are divided into a 70% training set and a 30% validation set. The proportion of minority class samples in the training sample set is adjusted, and the classifier is continuously trained to improve its ability to recognize minority class samples, so that sample augmentation can be achieved in subsequent steps based on its initial classification results.

[0143] Specifically, the classifier is tuned during training and classification by evaluating metrics to obtain a classifier with the maximum recall. Evaluation metrics include: confusion matrix results, overall precision, and recall.

[0144] The confusion matrix results include: true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN), which are used to determine the classification accuracy of several classifiers.

[0145] The formula for overall accuracy is:

[0146]

[0147] Overall Accuracy refers to the overall accuracy.

[0148] The formula for recall rate is:

[0149]

[0150] Here, Recall is the recall rate.

[0151] S5: Perform sample augmentation on the minority classification results among several first classification results to obtain a second sample set;

[0152] In one alternative implementation, such as Figure 3 As shown, S5 includes:

[0153] S5.1: The minority classification results from several first-class classification results are compared with the forest layer data to determine the overlapping area. The overlapping area is:

[0154] A overlap =A s ∩A f (20)

[0155] Among them, A overlap A represents a small number of overlapping areas between classification results and forest layer data; s A region with a small number of classification results; f This refers to the region of the forest layer data.

[0156] S5.2: Create a buffer centered on the few classification results located in the overlapping region, generate new sample points in the buffer to achieve sample augmentation, and obtain a second sample set based on the augmented samples;

[0157] The formula for the buffer is:

[0158] B i ={(x,y)|(xx) i ) 2 +(yy i ) 2 ≤r 2}; (twenty one)

[0159] Where (x,y) is any point within the buffer; p i For the minority classification result, its coordinates are (x i ,y i );B i To classify the minority results p i A buffer zone with center r and radius r.

[0160] The formula for the new sample point is:

[0161]

[0162] Where, p new For new sample points generated within the buffer; (x j ,y j ) represents the coordinates of the new sample point generated within the buffer; n new This represents the number of new sample points generated within the buffer.

[0163] The principle is as follows: by comparing the minority classification results with the forest layer data, a map of the minority class distribution is obtained. Then, based on the minority class distribution map, a buffer is combined to obtain a base layer for generating new sample points. New sample points are generated within this base layer, meaning that the generated enhanced sample points are all located within a reasonable forest cover area.

[0164] S6: Train several first classifiers based on the second sample set, and obtain several second classifiers accordingly; classify the preprocessed multi-source remote sensing data using the several second classifiers, and obtain several second classification results; then perform decision fusion on the several second classification results to obtain the final classification result.

[0165] In one alternative implementation, such as Figure 4 As shown, S6 includes:

[0166] S6.1: Train several first classifiers based on the second sample set, and obtain several second classifiers accordingly;

[0167] For example, using a second sample set as input, three classifiers are trained: a Random Forest (RF) classifier, a Decision Tree (DT) classifier, and a Support Vector Machine (SVM) classifier. Specifically, in the Random Forest classifier, the final classification result is generated by constructing multiple decision trees and taking the average or majority vote of their predictions. Each decision tree is trained with a different subset of samples and features, thus reducing the overfitting problem of a single decision tree and improving the model's generalization ability. Similarly, in the Decision Tree classifier, the final classification result is generated by constructing multiple decision trees and taking the average or majority vote of their predictions. Each decision tree is trained with a different subset of samples and features, thus reducing the overfitting problem of a single decision tree and improving the model's generalization ability. In the Support Vector Machine classifier, the optimal hyperplane is found to separate samples of different classes, and the distance from the hyperplane to the nearest sample point is maximized, making it particularly effective for classification problems in high-dimensional spaces.

[0168] S6.2: Classify the preprocessed multi-source remote sensing data according to several second classifiers to obtain several classification results;

[0169] S6.3: Perform decision fusion on several secondary classification results to obtain the final classification result, which includes:

[0170]

[0171] in, The final classification result is represented by z, where z is the pixel value and mode is the function that calculates the category that appears most frequently among the three classification results, i.e., the maximum voting method. and The results are the classification results of several second classifiers, in sequence.

[0172] In addition to random forest classifiers, decision tree classifiers, and support vector machine classifiers, other machine learning algorithms or classifiers can be selected for classification, such as Extreme Gradient Boosting (XGBoost), deep learning models, and k-nearest neighbor classifiers. Decision fusion can also be performed based on other decision fusion strategies, such as weighted voting and adaptive enhancement methods.

[0173] The multi-source remote sensing extraction method for planted forests of the present invention is based on multi-source remote sensing data. It performs sample enhancement based on the first classification result of the trained first classifier and generates new sample points in the buffer of minority class sample points. This not only effectively increases the number of minority class samples, but also ensures the reasonable distribution of new sample points, improves the classifier's ability to identify minority class samples, and solves the problem of class imbalance.

[0174] The multi-source remote sensing extraction method for planted forests of the present invention utilizes the temporal characteristics of NDVI processing and extracts trend and seasonal change features in the time series through harmonic regression fitting. This not only enhances the classifier's ability to capture vegetation changes but also improves the classifier's sensitivity to long-term and periodic changes, providing support for obtaining more accurate classification results.

[0175] The multi-source remote sensing extraction method for planted forests of the present invention trains three different classifiers—random forest, decision tree, and support vector machine—and performs decision fusion to generate the final classification result. This fusion strategy can effectively integrate the advantages of different classifiers, overcome the bias problem that is prone to occur with a single classifier, and improve the overall accuracy, stability, and robustness of the final classification result.

[0176] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0177] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for multi-source remote sensing extraction of a plantation, characterized in that, The method comprises the following steps: acquiring multi-source remote sensing data; preprocessing the multi-source remote sensing data to obtain preprocessed multi-source remote sensing data; acquiring training samples and verification samples respectively, and screening the training samples according to the verification samples to obtain a first sample set; training a plurality of classifiers according to the first sample set, obtaining a plurality of first classifiers correspondingly, and classifying the first sample set through the plurality of first classifiers correspondingly to obtain a plurality of first classification results; performing sample enhancement on a minority of the plurality of first classification results to obtain a second sample set; wherein the sample enhancement comprises: performing coincidence judgment on the minority of the plurality of first classification results and the verification samples in the first sample set to obtain a coincidence region; creating a buffer zone centered on the minority of the plurality of first classification results in the coincidence region, generating new sample points in the buffer zone to realize sample enhancement, and obtaining a second sample set according to the enhanced samples; training the plurality of first classifiers according to the second sample set, obtaining a plurality of second classifiers correspondingly; classifying the preprocessed multi-source remote sensing data through the plurality of second classifiers correspondingly to obtain a plurality of second classification results, and performing decision fusion on the plurality of second classification results to obtain a final classification result.

2. The method according to claim 1, wherein, The method comprises the following steps:

3. The method according to claim 2, wherein, acquiring multi-source remote sensing data, including: acquiring Sentinel-2 time series images, Sentinel-1 time series images and Landsat time series images respectively. The method comprises the following steps: preprocessing the multi-source remote sensing data to obtain preprocessed multi-source remote sensing data, including: performing cloud removal processing on the Sentinel-2 time series images, the Sentinel-1 time series images and the Landsat time series images respectively to obtain multi-spectral band information; performing NDVI processing on the Landsat time series images after cloud removal to obtain a normalized vegetation index; performing harmonic regression fitting on the normalized vegetation index, and extracting statistical features of the fitted normalized vegetation index; wherein the statistical features include: mean, variance, standard deviation, coefficient of variation and modified variance; 4. The method according to claim 3, wherein, The multi-source remote sensing data further comprises: terrain feature information and radar image data information; multi-source feature integration is performed on the multi-spectral band information, the statistical features, the terrain feature information and the radar image data information to obtain an integrated result; ； wherein, NDVI is the normalized difference vegetation index; Rnir is the reflectance value in the near infrared band; Rred is the reflectance value in the red band.

5. The method according to claim 4, wherein, extracting the integrated result to obtain texture features. ； wherein is a set of normalized vegetation indices after fitting reconstruction; is the number of harmonics; is a time step, , is the number of observations in the time series; and are coefficients of trigonometric terms in the Fourier series, is the corresponding frequency; is a constant term.

6. The method according to claim 1, wherein, The formula of the NDVI processing is: The model of the harmonic regression fitting is: The method comprises the following steps: acquiring training samples and verification samples respectively, and screening the training samples according to the verification samples to obtain a first sample set, including: acquiring training samples and verification samples respectively, and the verification samples are forest layer data; comparing the training samples and the forest layer data, and screening out the training samples located in the forest layer data as reserved samples; performing label re-projection on the reserved samples to obtain a first sample set.

7. The method according to claim 1, wherein, According to the first sample set, a plurality of classifiers are trained respectively, a plurality of first classifiers are correspondingly obtained, and the first sample set is classified by the plurality of first classifiers respectively, and a first classification result is correspondingly obtained, including: A few minority class samples and majority class samples are selected from the first sample set as a training sample set respectively; According to the training sample set, the plurality of classifiers are trained respectively; wherein the plurality of classifiers include: a random forest classifier, a decision tree classifier and a support vector machine classifier; The proportion of minority class samples in the training sample set is adjusted, and the plurality of classifiers are retrained according to the adjusted training sample set until an evaluation index is met, and a plurality of first classifiers are obtained; wherein the evaluation index is used to evaluate the classification accuracy of the plurality of trained classifiers; According to the plurality of first classifiers, the first sample set is classified respectively, and a first classification result is correspondingly obtained.

8. The method according to claim 7, wherein, The evaluation index includes: a confusion matrix result, an overall accuracy and a recall rate; The confusion matrix result includes: true positive examples , false positive examples , false negative examples , and true negative examples , for judging classification accuracy of the several classifiers. The formula of the overall accuracy is: ； wherein is the overall accuracy; The formula of the recall rate is: ； wherein, is the recall rate.

9. The method according to claim 6, wherein, The minority classification results in the plurality of first classification results are sample enhanced to obtain a second sample set, including: The minority classification results in the plurality of first classification results are coincided with the forest layer data to obtain a coincidence area, and the coincidence area is: ； wherein, is a region of overlap of the minority classification result and the forest layer data; is a region of the minority classification result; is a region of the forest layer data; A buffer area is created with the minority classification results in the coincidence area as the center, new sample points are generated in the buffer area to realize sample enhancement, and a second sample set is obtained according to the enhanced samples; The formula of the buffer area is: ； wherein, is the buffer region for any point within the buffer region; is the minority classification result, whose coordinates are ; is the buffer region with the center at the minority classification result and the radius of . The formula of the new sample point is: ； wherein, is a new sample point generated within the buffer; is a coordinate of the new sample point generated within the buffer; is a number of new sample points generated within the buffer.

10. The method according to claim 1, wherein, According to the second sample set, the plurality of first classifiers are trained respectively, and a plurality of second classifiers are correspondingly obtained, including: According to the second sample set, the plurality of first classifiers are trained respectively, and the plurality of second classifiers are correspondingly obtained; According to the plurality of second classifiers, the preprocessed multi-source remote sensing data is classified, and a plurality of second classification results are correspondingly obtained; The plurality of second classification results are decision fused to obtain a final classification result, and the final classification result includes: ； wherein, is the final classification result; is a pixel point; is a function for calculating the class with the highest occurrence frequency among the three classification results; , and are the classification results of the plurality of second classifiers in turn.

Images