Cross-period label sample automatic generation method

Abstract

The invention discloses a cross-period label sample automatic generation method, and aims to solve the problems of high sample labeling cost and insufficient model generalization ability in the prior art. The method comprises the following steps: acquiring and preprocessing remote sensing images in different periods; extracting space-time spectrum feature information by using methods such as wavelet transform, PCA, ICA and HSV; extracting ground feature information through ground feature indexes such as NDVI, NDWI, NDBI and BSI; constructing a constant ground feature index (CFI) formula, and quantifying the difference of the characteristic wave bands in the two-stage images; based on the CFI result, LISA spatial clustering analysis is carried out, and an LISA aggregation graph is generated; extracting a constant ground feature region through coupling analysis of CFI and LISA; and finally, superposing the constant ground object area with the first-stage label data to generate a cross-stage label sample data set. According to the method, the labeling cost is remarkably reduced, the sample quality and the model generalization ability are improved, and the method is suitable for fusion of various ground feature types and multi-source data and has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to the field of sample label generation, and particularly to an automatic generation method for cross-period label samples for dynamic monitoring oriented to deep learning. Background Art

[0002] In dynamic monitoring tasks driven by deep learning, the quality and scale of sample data directly determine the performance ceiling of the model. However, the sample production process faces many challenges. Especially in cross-period and cross-scenario dynamic monitoring tasks, the acquisition, annotation, and expansion of samples have become the key bottlenecks restricting the generalization ability of the model.

[0003] Existing sample production methods have many limitations. The commonly used transfer learning method in dynamic monitoring has a strong dependence on samples. Transfer learning can reduce the demand for labeled data in the target domain by reusing pre-trained models (such as ResNet trained on the ImageNet dataset). However, in dynamic monitoring tasks, the spectral, texture, and other features of the target scene are significantly different from those of the source domain, resulting in the need for a large amount of fine-tuning of the model. Research shows that the classification accuracy of directly transferred models in new scenarios is generally lower than 70%, and they are less robust to changes in lighting and seasons. In addition, transfer learning still requires a certain number of labeled samples in the target domain for fine-tuning, and the acquisition cost of cross-period label samples in dynamic monitoring tasks is relatively high.

[0004] Generally, to obtain high-precision dynamic monitoring results, the method of manually delineating multi-period label samples is adopted. By manually annotating two-phase images, a large amount of manpower is required for repeated annotation. Taking high-resolution images as an example, it takes several hours to delineate the target area of a single image, making it difficult to meet the timeliness requirements of large-scale monitoring.

[0005] Existing dynamic monitoring also uses end-to-end change detection models, which directly input two-phase images and change labels and output change models (such as Siamese networks, CDNet). They rely on accurate spatio-temporal registration and radiometric correction, while actual images often have registration deviations due to noise interference (such as cloud occlusion, sensor errors). Such models have a large number of parameters (for example, the U-Net++ variant requires more than 50M parameters), and face the trade-off between computing resources and inference speed during actual deployment. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide an automatic generation method for cross-period label samples for dynamic monitoring oriented to deep learning, so as to solve the contradiction between the annotation cost and the model generalization ability in existing deep learning dynamic monitoring, as well as the problem of insufficient utilization of label features.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A method for automatically generating cross - period label samples, comprising:

[0009] Step 1: Obtain the original remote sensing images of two periods containing the target area and perform pre - processing, namely the first - period remote sensing image and the second - period remote sensing image. Among them, the ground object label data of the first - period remote sensing image is known, and the original remote sensing image includes at least the data of three bands: R, G, and B.

[0010] Step 2: Extract spatio - temporal spectral features from the pre - processed first - period remote sensing image and second - period remote sensing image to generate spatio - temporal spectral feature information corresponding to the first - period remote sensing image and the second - period remote sensing image.

[0011] Step 3: Extract ground object index features from the pre - processed first - period remote sensing image and second - period remote sensing image to generate ground object index feature information corresponding to the first - period remote sensing image and the second - period remote sensing image.

[0012] Step 4: Construct a calculation formula for the constant ground object index CFI. Among them, the constant ground object index CFI is related to the target feature band data of two periods that can best reflect the change of ground object difference in the target area.

[0013] Step 5: According to the specific situation of the target area, extract the target feature band data from the spatio - temporal spectral feature information and the ground object index feature information, and use the selected target feature band data to calculate the CFI result data of the target area.

[0014] Step 6: Perform LISA spatial clustering on the CFI result data of the target area to generate LISA spatial clustering results. The LISA spatial clustering results include high - high clustering areas, high - low clustering areas, low - high clustering areas, low - low clustering areas, and insignificant areas.

[0015] Step 7: Perform coupling analysis on the CFI result data and the LISA spatial clustering results of the target area to determine the constant ground object area of the target area.

[0016] Step 8: According to the ground object label data of the first - period remote sensing image, determine the label data of the constant ground object area in the target area of the second - period remote sensing image, thereby generating a cross - period label sample data set.

[0017] Further, the pre - processing includes one or a combination of geometric correction, radiometric correction, atmospheric correction, cropping, etc.

[0018] Further, the method for extracting spatio - temporal spectral features includes one or a combination of wavelet transform, PCA, ICA, HSV transform methods.

[0019] Further, the method for extracting feature indices of ground objects includes extracting using one or a combination of multiple indices such as NDVI, NDWI, NDBI, BSI, EVI, and SAVI.

[0020] Further, the calculation formula for the constant feature index of ground objects CFI is:

[0021]

[0022] Among them, Te1 represents the normalized feature band data of the first period; Te2 represents the normalized feature band data of the first period; represents the α feature band data of the first period; represents the β feature band data of the first period; represents the α feature band data of the second period; represents the β feature band data of the second period; and are the target feature band data.

[0023] Further, step 7 specifically includes:

[0024] Set the CFI threshold range [a, b];

[0025] According to the CFI result data of the target area and the CFI threshold range, determine the basic constant ground object information;

[0026] Couple the LISA spatial clustering result and the basic constant ground object information to generate the final constant ground object information, so as to determine the constant ground object area of the final target area.

[0027] Further, according to the CFI result data of the target area and the CFI threshold range, determining the basic constant ground object information includes:

[0028] Preliminarily determine the area where a ≤ CFI ≤ b in the target area as the constant ground object area;

[0029] Preliminarily determine the area where CFI b in the target area as the changing area.

[0030] Further, the method specifically includes:

[0031] Determine the areas that simultaneously satisfy CFI < a and the LISA clustering result is low-low clustering, and the areas that simultaneously satisfy CFI > b and the LISA clustering result is high-high clustering as the changing areas, and determine the remaining areas as the constant areas of the final target area.

[0032] Further, for the CFI threshold range and the combination method of CFI and the spatial clustering result, it can be adjusted according to the specific situation of the target area.

[0033] The beneficial effects of the present invention are as follows:

[0034] First, it significantly reduces the annotation cost: By constructing a Constant Feature Index (CFI) and combining it with LISA spatial clustering analysis, the present invention can automatically generate cross - period pseudo - label data based on a single - period annotation sample, avoiding the high cost of manual annotation in each period in traditional methods;

[0035] Second, it improves the sample quality and the generalization ability of the model: The present invention utilizes the temporal stability of spectral features. Through spectral similarity matching and spatial clustering analysis, it effectively extracts the constant feature regions in cross - period images and generates high - quality pseudo - label data. These data can significantly enhance the generalization ability of deep learning models, making them perform more stably in dynamic monitoring tasks across different periods and scenarios;

[0036] Third, it has strong adaptability and is applicable to various feature types: By combining multiple spatio - spectral feature extraction methods (such as PCA, ICA, HSV, etc.) and feature index analysis (such as NDVI, NDWI, NDBI, BSI, etc.), the present invention can adapt to the feature extraction requirements of different feature types (such as vegetation, water bodies, buildings, bare soil, etc.) and has a wide range of application scenarios;

[0037] Fourth, it improves the change detection accuracy: Through the coupled analysis of the CFI index and LISA spatial clustering, the present invention can accurately distinguish the change regions from the constant feature regions and effectively reduce noise interference.

[0038] Fifth, it supports multi - source data fusion: The present invention is not only applicable to RGB three - band images, but can also be extended to multi - spectral and hyperspectral image data. By fusing multi - source data, it further improves the accuracy and robustness of feature extraction;

[0039] Sixth, it has a high degree of automation and intelligence: The present invention realizes the full automation from image pre - processing, feature extraction to sample generation, reduces manual intervention, improves the intelligence level of dynamic monitoring, and provides efficient technical support for large - scale remote sensing image analysis.

[0040] In summary, by innovatively combining spectral feature analysis, constant feature index construction, and spatial clustering methods, the present invention solves the problems of high cost, low efficiency, and poor quality in sample production for deep - learning - based dynamic monitoring, and provides an efficient, accurate, and highly adaptable technical solution for remote sensing image dynamic monitoring.

[0041] Other advantages, objects, and features of the present invention will be set forth to some extent in the following description, and to some extent, will be apparent to those skilled in the art based on the examination of the following, or can be taught from the practice of the present invention. The objects and other advantages of the present invention can be achieved and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings, where:

[0043] Figure 1 is a schematic flowchart of an automatic generation method for cross - period label samples shown according to an embodiment of the present application;

[0044] Figure 2 is a specific technical roadmap for the implementation of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The following will refer to the accompanying drawings to describe the preferred embodiments of the present invention in detail. It should be understood that the preferred embodiments are only for illustrating the present invention, rather than limiting the protection scope of the present invention.

[0046] Combined with Figure 1 and Figure 2 , the automatic generation method for cross - period label samples for deep - learning dynamic monitoring includes the following steps:

[0047] Step 1: Obtain and pre - process the original remote - sensing images of two periods containing the target area, namely the first - period remote - sensing image and the second - period remote - sensing image. Among them, the ground - object label data of the first - period remote - sensing image is known, and the original remote - sensing image includes at least data of three bands, namely R, G, and B. The first period can be earlier or later than the second period;

[0048] Step 2: Extract spatio - temporal spectral features from the pre - processed first - period and second - period remote - sensing images to generate spatio - temporal spectral feature information corresponding to the first - period and second - period remote - sensing images;

[0049] Step 3: Extract ground - object index features from the pre - processed first - period and second - period remote - sensing images to generate ground - object index feature information corresponding to the first - period and second - period remote - sensing images;

[0050] Step 4: Construct a calculation formula for the constant ground - object index CFI, where the constant ground - object index CFI is related to the target feature band data of two periods that can best reflect the change in ground - object difference in the target area;

[0051] Step 5: According to the specific situation of the target area, extract the target characteristic band data from the spatiotemporal spectrum characteristic information and the ground feature index characteristic information, and use the selected target characteristic band data to calculate the CFI result data of the target area;

[0052] Step 6: Perform LISA spatial clustering on the CFI result data of the target area to generate LISA spatial clustering results. The LISA spatial clustering results include high-high clustering areas, high-low clustering areas, low-high clustering areas, low-low clustering areas, and insignificant areas.

[0053] Step 7: Conduct coupling analysis on the CFI result data and LISA spatial clustering results of the target area to determine the constant feature area in the target area;

[0054] Step 8: Based on the object label data of the first phase remote sensing image, determine the label data of the constant object area in the target area of the second phase remote sensing image, thereby generating a cross-period label sample data set.

[0055] In step 1, the raw remote sensing image data can be from an unmanned aerial vehicle (UAV), satellite, or other remote sensing platform, and the resolution can be selected based on actual needs (e.g., 0.2 m, 0.5 m, 1 m, etc.). The RGB three-band data ensures that basic spectral feature analysis operations can be performed on it. In other embodiments, the raw remote sensing image data can also be multispectral or hyperspectral image data.

[0056] The preprocessing may include one or more combinations of geometric correction, radiation correction, atmospheric correction, clipping, etc.

[0057] Geometric correction eliminates geometric distortion caused by factors such as sensor attitude and terrain undulation, ensuring accurate spatial positioning of the image. Radiometric correction eliminates differences in sensor response and the effects of lighting conditions, ensuring that the image's radiometric values reflect the true reflectivity of the ground objects. Atmospheric correction removes the effects of atmospheric scattering and absorption, restoring the true spectral characteristics of the ground objects. Cropping, based on the scope of the study area, cuts out the image data of the target area, reducing data volume and improving processing efficiency. The preprocessed image data is then used for subsequent feature extraction and analysis.

[0058] In step 2, the method for extracting spatiotemporal spectral features includes one or more combinations of wavelet transform, principal component analysis (PCA), independent component analysis (ICA), and HSV (hue, saturation, value) transformation methods.

[0059] Wavelet transform extracts the texture and edge features of images through multi-scale decomposition, and is suitable for capturing local detail information of ground objects. For example, it can extract the outlines of buildings, the distribution of vegetation, etc.

[0060] PCA reduces the dimensionality of multi-band data through linear transformation, reduces the number of bands, extracts the main change information, reduces data redundancy, and enhances specific ground objects (such as water bodies, bare soil, etc.).

[0061] ICA separates the independent components in remote sensing images through the statistical independence hypothesis. Each component represents a statistically independent source signal (such as separating independent components such as vegetation, soil, and water bodies), and is suitable for extracting target features in mixed signals.

[0062] HSV converts RGB images into the HSV color space, separates color information and brightness information, and facilitates ground object classification and target detection.

[0063] The result data (i.e., spatio-temporal spectral feature information) extracted according to the above methods includes multi-band information, which can reflect the spatio-temporal and spectral characteristics of images and provides a basis for the construction of subsequent constant ground object indices.

[0064] In step 3, the methods for extracting ground object index features include using one or a combination of indices such as the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Water Index (NDWI), the Normalized Difference Built-up Index (NDBI), the Bare Soil Index (BSI), the Enhanced Vegetation Index (EVI), and the Soil-Adjusted Vegetation Index (SAVI) for extraction.

[0065] NDVI is used to extract vegetation information and is expressed by the formula:

[0066]

[0067] Among them, NIR represents the near-infrared band data, and Red represents the red band data.

[0068] NDWI is used to extract water body information and is expressed by the formula:

[0069]

[0070] Among them, Green represents the green band data.

[0071] The NDBI is used to extract building information, which is expressed by the formula:

[0072]

[0073] Among them, SWIR represents the data of the near-infrared band.

[0074] The BSI is used to extract bare soil information, which is expressed by the formula:

[0075]

[0076] The EVI is used to evaluate the vegetation growth situation, which can reduce the influence of land cover. Especially in the hyperspectral case, it can better reflect the vegetation growth situation, and is expressed by the formula:

[0077]

[0078] Among them, Blue represents the data of the blue band.

[0079] The SAVI aims to reduce the influence of the soil background on the vegetation index by adjusting the parameter L, so as to improve the accuracy of vegetation condition assessment, and is expressed by the formula:

[0080]

[0081] L is a parameter that changes with the vegetation density, and its value range is from 0 to 1. When the vegetation coverage is very high, L is 0; when the vegetation coverage is very low, L is 1. If L = 0, then SAVI = NDVI.

[0082] The result data (i.e., the feature information of the ground object index) extracted from the above ground object indexes includes multi-band information, which can reflect the spectral characteristics of different ground objects and provide support for the construction of the constant ground object index.

[0083] In step 4, the calculation formula of the constant ground object index CFI is:

[0084]

[0085] Among them, Te1 represents the normalized characteristic band data of the first period; Te2 represents the normalized characteristic band data of the first period; represents the α characteristic band data of the first period; represents the β characteristic band data of the first period; represents the α characteristic band data of the second period; represents the β characteristic band data of the second period; and are the target characteristic band data.

[0086] By quantifying the differences in characteristic bands between two-phase images, the CFI can effectively reflect the changes in ground objects.

[0087] In step 5, according to the specific situation of the target area, the target characteristic band data is extracted from the spatio-temporal spectral characteristic information and the ground object index characteristic information, and the CFI result data of the target area is calculated by using the selected target characteristic band data. For different target areas, the types of ground objects existing in the area may be different, and moreover, the changes in ground objects in two periods may also be different. Therefore, when selecting the target characteristic band data, the selected target band data will also be different. That is to say, in the calculation of CFI, the selection of the α band and the β band is related to the specific situation of the target area.

[0088] For a target area, the CFI result is a spatial distribution map, which can intuitively display the changing area and the constant area of the ground object. Through the CFI result, the changes in the ground object can be preliminarily judged, providing a basis for subsequent spatial clustering analysis.

[0089] In step 6, the Anselin Local Moran's I method can be used to perform spatial clustering analysis on the CFI result to generate a LISA clustering map. Anselin Local Moran's I (local Moran's index) is a method of local spatial autocorrelation analysis proposed by the spatial statistician Luc Anselin in 1995. It is a localized version of the global Moran's I, used to identify the spatial association patterns in local areas of spatial data. Through Anselin Local Moran's I analysis, a LISA map can be obtained, and the LISA map can distinguish the following four spatial relationships: high-high clustering (high CFI value areas are clustered), low-low clustering (low CFI value areas are clustered), high-low clustering (high CFI value areas are surrounded by low CFI value areas), low-high clustering (low CFI value areas are surrounded by high CFI value areas), and insignificant distribution.

[0090] The LISA clustering result can further refine the spatial distribution characteristics of the ground object changes and provide support for the extraction of the constant ground object area.

[0091] In some embodiments, step 7 may specifically include:

[0092] First, set the CFI threshold range [a, b], and this CFI threshold range can be adjusted according to the specific situation of the target area, and then further divide according to the threshold;

[0093] Next, based on the CFI result data and the CFI threshold range of the target area, the basic constant feature information is determined. The area where a ≤ CFI ≤ b in the target area can be preliminarily determined as the constant feature area, and the area where CFI b in the target area can be preliminarily determined as the changing area. By threshold division, the basic constant feature information is extracted, which can effectively distinguish the changing area and the constant area, providing a basis for subsequent coupling analysis.

[0094] According to the LISA spatial clustering result and the basic constant feature information, the final constant feature information is generated, so as to determine the constant feature area of the final target area, that is, the areas that simultaneously satisfy CFI b and the LISA clustering result is high-high clustering are determined as the changing areas, and the remaining areas are determined as the constant areas of the final target area.

[0095] Through the coupling analysis of the CFI result data and the LISA spatial clustering result, the extraction accuracy of the constant feature area can be further improved, and the noise interference can be reduced. The finally generated cross-temporal label data can significantly reduce the annotation cost, improve the sample quality and the model generalization ability, providing efficient and accurate technical support for deep learning dynamic monitoring.

[0096] In step 8, the determined constant feature area can be spatially overlaid with the first-phase label data (i.e., the feature label data of the first-phase remote sensing image), and the intersection part is taken to generate cross-temporal label data. The generated label data is cut to obtain label sample data sets of different periods for the training and verification of the deep learning model. Through this step, the annotation cost can be significantly reduced, and the sample quality and the model generalization ability can be improved.

[0097] Embodiment 1

[0098] Step 1: Obtain two-phase high-resolution image data including the target area, including the RGB three bands and the near-infrared band, and perform preprocessing on the data, including geometric correction, radiometric correction, atmospheric correction, and cropping, to obtain the preprocessed two-phase image data;

[0099] Step 2: Perform PCA, ICA, and HSV transformations on the preprocessed first-phase remote sensing image and the second-phase remote sensing image, extract spatio-spectral feature information, and generate PCA, ICA, and HSV result data corresponding to the first-phase remote sensing image and the second-phase remote sensing image;

[0100] Step 3: Extract the feature index of the preprocessed first-phase remote sensing image and the second-phase remote sensing image, calculate the NDVI, NDWI, BSI, and EVI indices, and generate NDVI, NDWI, BSI, and EVI result data corresponding to the first-phase remote sensing image and the second-phase remote sensing image;

[0101] Step 4: Construct the calculation formula of the Constant Feature Index (CFI). The CFI is related to the target feature band data that can best reflect the change of feature differences in the target area in two periods. The calculation formula of the CFI is as follows:

[0102]

[0103] where Te1 represents the normalized feature band data in the first period; Te2 represents the normalized feature band data in the first period; represents the α feature band data in the first period; represents the β feature band data in the first period; represents the α feature band data in the second period; represents the β feature band data in the second period; and are the target feature band data;

[0104] Step 5: According to the specific situation of the target area, extract the target feature band data from the PCA, ICA, HSV result data and NDVI, NDWI, BSI, EVI results. After selection, the α feature band data is the NDVI result data, and the β feature band data is the NDWI result data. Substitute them into the formula to calculate the CFI result data of the target area;

[0105] Step 6: Conduct LISA spatial clustering on the CFI result data of the target area to generate LISA spatial clustering results, which include high-high clustering areas, high-low clustering areas, low-high clustering areas, low-low clustering areas and insignificant areas;

[0106] Step 7: Conduct coupling analysis on the CFI result data and the LISA spatial clustering results of the target area. Set the CFI threshold range as [-0.1, 0.2]. Determine the changing areas as those that simultaneously satisfy CFI < -0.1 and the LISA clustering result is low-low clustering, and those that simultaneously satisfy CFI > 0.2 and the LISA clustering result is high-high clustering. Determine the remaining areas as the constant areas of the final target area, and extract the constant areas;

[0107] Step 8: According to the feature label data of the first-phase remote sensing image, spatially overlay the constant area obtained in Step 7 to obtain the label data of the constant feature area in the target area of the second-phase remote sensing image, thereby generating a cross-period label sample dataset.

[0108] In this example, a method for extracting cross - period label samples based on multi - temporal high - resolution remote sensing images is elaborated in detail. This method constructs a constant feature index CFI to quantify the temporal change differences of ground objects by fusing multi - source features (including PCA / ICA / HSV transformation features and ground object index features such as NDVI / NDWI / BSI / EVI), and combines LISA spatial clustering analysis to accurately separate the changed areas from the constant areas. Finally, through spatial overlay analysis, the ground object labels of the first period are migrated to the constant areas of the second period, automatically generating a cross - period label sample dataset. This technical solution effectively solves the problem of automatic extraction of cross - temporal samples of high - resolution remote sensing images, provides reliable sample support for remote sensing applications such as change detection and ground object classification, and is especially suitable for large - scale and long - time - series remote sensing monitoring tasks. The entire process has technical characteristics such as comprehensive feature fusion, scientific threshold setting, and strong spatial correlation, and has important application value in fields such as land monitoring and ecological environment assessment.

[0109] Example 2

[0110] This example is an explanatory note based on Example 1. Specifically:

[0111] Step 1: Obtain two - period UAV image data containing the target area, including RGB three - band data, and pre - process the data, including geometric correction, radiometric correction, and cropping, to obtain the pre - processed two - period image data;

[0112] Step 2: Perform PCA, ICA, and HSV transformations on the pre - processed first - period and second - period remote sensing images, extract spatio - spectral feature information, and generate PCA, ICA, and HSV result data corresponding to the first - period and second - period remote sensing images;

[0113] Step 3: Delete and do not calculate the relevant ground object index data;

[0114] Step 4 is the same as in Example 1;

[0115] Step 5: According to the specific situation of the target area, extract the target feature band data from the PCA, ICA, and HSV result data. After selection, the α feature band data is the first - band data of the HSV result data, and the β feature band data is the second - band data of the HSV result data. Substitute them into the formula to calculate the CFI result data of the target area;

[0116] Step 6 is the same as in Example 1;

[0117] Step 7: Couple the CFI result data and LISA spatial clustering results of the target area and analyze them. Set the CFI threshold range to [-0.2, 0.6], and identify the areas that satisfy both CFI < -0.2 and LISA clustering results as low-low clusters, and the areas that satisfy both CFI > 0.6 and LISA clustering results as high-high clusters as changing areas. The remaining areas are identified as constant areas of the final target area, and the constant areas are extracted.

[0118] Step 8 is the same as Example 1.

[0119] In this example, we used drone imagery data, using only RGB three-band data. It's important to note that correlation index analysis cannot be performed using only RGB three-band data, so step 3 is omitted in this example, and only step 2, the correlation transformation analysis, is performed.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for automatically generating cross-period label samples, characterized in that: Including: Step 1: Obtain the original remote sensing images of two periods including the target area and perform preprocessing, namely the first-period remote sensing image and the second-period remote sensing image. Among them, the ground object label data of the first-period remote sensing image is known, and the original remote sensing image includes at least the data of three bands of R, G, and B; Step 2: Extract the spatio-temporal spectral features of the preprocessed first-period remote sensing image and second-period remote sensing image to generate the spatio-temporal spectral feature information corresponding to the first-period remote sensing image and the second-period remote sensing image; Step 3: Extract the ground object index features of the preprocessed first-period remote sensing image and second-period remote sensing image to generate the ground object index feature information corresponding to the first-period remote sensing image and the second-period remote sensing image; Step 4: Construct the calculation formula of the constant ground object index CFI. Among them, the constant ground object index CFI is related to the target feature band data that can best reflect the ground object difference change situation in two periods; Step 5: According to the specific situation of the target area, extract the target feature band data from the spatio-temporal spectral feature information and the ground object index feature information, and use the selected target feature band data to calculate the CFI result data of the target area; Step 6: Perform LISA spatial clustering on the CFI result data of the target area to generate the LISA spatial clustering result, and the LISA spatial clustering result includes high-high clustering areas, high-low clustering areas, low-high clustering areas, low-low clustering areas, and insignificant areas; Step 7: Determine the constant ground object area of the target area according to the CFI result data and the LISA spatial clustering result of the target area; Step 8: According to the ground object label data of the first-period remote sensing image, determine the label data of the constant ground object area in the target area of the second-period remote sensing image, so as to generate a cross-period label sample data set.

2. The automatic generation method of an intertemporal label sample according to claim 1, wherein The preprocessing includes one or a combination of geometric correction, radiometric correction, atmospheric correction, cropping, etc.

3. The automatic generation method of an intertemporal label sample according to claim 1, characterized in that, The methods for extracting spatio-temporal spectral features include one or a combination of wavelet transform, PCA, ICA, HSV transform methods, etc.

4. The method for automatically generating label samples across time periods according to claim 1, characterized in that: The methods for extracting ground object index features include extraction using one or a combination of indexes such as NDVI, NDWI, NDBI, BSI, EVI, SAVI, etc.

5. The automatic generation method of an intertemporal label sample according to claim 1, characterized in that The calculation formula of the constant ground object index CFI is: Among them, Te1 represents the normalized characteristic band data of the first period; Te2 represents the normalized characteristic band data of the first period; represents the α characteristic band data of the first period; represents the beta characteristic band data of the first period; represents the α characteristic band data of the second period; represents the β characteristic band data of the second period; and This is the target characteristic band data.

6. The automatic generation method of an intertemporal label sample according to claim 1, characterized in that Step 7 specifically includes: Set the CFI threshold range [a, b]; Determine the basic constant ground object information according to the CFI result data and the CFI threshold range of the target area; Couple the LISA spatial clustering result and the basic constant ground object information to generate the final constant ground object information, so as to determine the constant ground object area of the final target area.

7. The method for automatically generating cross-period label samples according to claim 6, characterized in that: Determine the basic constant ground object information according to the CFI result data and the CFI threshold range of the target area, including: Preliminarily determine the area where a ≤ CFI ≤ b in the target area as the constant ground object area; Preliminarily determine the area where CFI b in the target area as the changing area.

8. The method for automatically generating cross-period label samples according to claim 7, characterized in that: The method specifically includes: Regions that simultaneously satisfy CFI < a and have a low-low clustering result in the LISA clustering, as well as regions that simultaneously satisfy CFI > b and have a high-high clustering result in the LISA clustering, are determined as changing regions, and the remaining regions are determined as the constant regions of the final target regions.

9. The automatic generation method of an intertemporal label sample according to claim 6, wherein, The CFI threshold range and the combination method of CFI and spatial clustering results can be adjusted according to the specific situation of the target region.

Images