Remote sensing land cover classification method, apparatus and device based on transfer learning, and storage medium

Abstract

Provided in the present application are a remote sensing land cover classification method, apparatus and device based on transfer learning, and a storage medium. The method comprises: acquiring a source period image, a classification result label corresponding to the source period image, and a plurality of target period images; separately calculating a spectral slope vector corresponding to the source period image and a spectral slope vector corresponding to each target period image; on the basis of the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image, calculating a change intensity of each target period image with respect to the source period image; on the basis of the change intensities and the classification result label, determining an initial prediction sample for supervised classification; performing purification processing on the initial prediction sample, so as to obtain a trusted sample for supervised classification; and on the basis of the trusted sample, performing supervised classification on the plurality of target period images, so as to obtain a target land cover product. Therefore, the effect of improving the temporal resolution, classification efficiency and accuracy of remote sensing land cover classification products can be achieved.

Description

Remote sensing land cover classification methods, devices, equipment, and storage media based on transfer learning

[0001] This application claims priority to Chinese Patent Application No. 202510080996.X, filed on January 20, 2025, entitled “Method, Apparatus, Device and Storage Medium for Remote Sensing Land Cover Classification under Transfer Learning”, the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the field of remote sensing information processing technology, and in particular to a method, apparatus, equipment and storage medium for remote sensing land cover classification based on transfer learning. Background Technology

[0003] Land cover classification is one of the important applications of remote sensing technology. It aims to classify each pixel in a remote sensing image into a specific land cover type. Analyzing the classification and changes of land use / cover in watersheds is the foundation for adaptive watershed management. Therefore, developing more efficient remote sensing land cover classification methods has become a promising direction for application.

[0004] In existing technologies, land cover classification methods can be divided into two types: supervised classification and unsupervised classification. Supervised classification is suitable for situations where prior knowledge is known. It involves a large amount of sample labeling work, but it has high accuracy and a wide range of applications, making it the most widely used method. Unsupervised classification, on the other hand, does not require prior knowledge and performs classification directly.

[0005] However, existing supervised classification methods require manual labeling of different types of land cover samples, which greatly increases time and labor costs. Unsupervised classification methods, on the other hand, cannot eliminate interference from heterogeneous objects with the same spectral density and heterogeneous objects with the same spectral density, resulting in reduced accuracy. These methods suffer from technical problems such as low temporal resolution, low classification efficiency, and poor classification accuracy in remote sensing land cover classification products. Summary of the Invention

[0006] This application provides a method, apparatus, equipment, and storage medium for remote sensing land cover classification using transfer learning, in order to improve the temporal resolution, classification efficiency, and accuracy of remote sensing land cover classification products.

[0007] Firstly, this application provides a remote sensing land cover classification method based on transfer learning, including:

[0008] Acquire source period images, corresponding classification result labels for source period images, and multiple target period images;

[0009] Calculate the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image respectively;

[0010] Based on the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image, calculate the change intensity between each target period image and the source period image.

[0011] Based on the intensity of change and the classification result labels, the initial prediction samples for supervised classification are determined;

[0012] The initial prediction samples are purified to obtain reliable samples for supervised classification;

[0013] Based on reliable samples, supervised classification is performed on images from multiple target periods to obtain target land cover products.

[0014] In one possible implementation, the initial prediction samples are purified to obtain reliable samples for supervised classification, including:

[0015] Based on the clustering characteristics of the spatial distribution of ground features, broken tags in the initial predicted samples are removed to obtain the first purified sample;

[0016] The first purified sample was characterized to retain the tags whose intensity of change was within the range of change intensity, thus obtaining the second purified sample;

[0017] The Markov random field model and information entropy theory are used to process the second purified sample to obtain a reliable sample for supervised classification.

[0018] In one possible implementation, a Markov random field model and information entropy theory are used to process the second purified sample to obtain a reliable sample for supervised classification, including:

[0019] The second purified sample was used as a label to pre-classify images of multiple target periods using a support vector machine model to obtain the pre-classification posterior probability and pre-classification results.

[0020] Based on the pre-classification results, a Markov random field model is introduced to output a normalized total energy image;

[0021] For the pre-classification posterior probability, information entropy theory is introduced to calculate the entropy value map;

[0022] Based on the normalized total energy image and entropy map, a weighted map is generated, and a reliable region is determined based on the weighted map;

[0023] The trusted region is spatially overlaid with the second purified sample to obtain a trusted sample for supervised classification.

[0024] In one possible implementation, initial prediction samples for supervised classification are determined based on the intensity of change and the classification result label, including:

[0025] Based on the classification results labels and the intensity of change, multiple intensity of change zones were determined;

[0026] For multiple zones with varying intensity, multiple zone thresholds are determined;

[0027] Based on multiple partition thresholds and the intensity of change, the unchanged areas are identified;

[0028] The unchanged regions and classification result labels are overlaid as output to obtain the initial prediction samples for supervised classification.

[0029] In one possible implementation, the first purified sample is characterized to retain tags whose intensity of change is within the range of intensity change, resulting in a second purified sample, comprising:

[0030] Based on the intensity of change and the unchanged area, determine the mean and variance of the intensity of change corresponding to each intensity of change zone;

[0031] Based on the mean and variance, determine the range of change intensity corresponding to each change intensity zone;

[0032] Based on the range of change intensity, the first purified sample is screened to retain the tags whose change intensity is within the range of change intensity, thus obtaining the second purified sample.

[0033] In one possible implementation, the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image are calculated, including:

[0034] Calculate the spectral slope corresponding to the source period image and the spectral slope corresponding to each target period image respectively;

[0035] Based on the spectral slope of the source period image and the spectral slope of each target period image, calculate the spectral slope vector of the source period image and the spectral slope vector of each target period image.

[0036] Secondly, this application provides a remote sensing land cover classification device based on transfer learning, comprising:

[0037] The acquisition module is used to acquire source period images, classification result labels, and multiple target period images;

[0038] The first calculation module is used to calculate the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image, respectively.

[0039] The second calculation module is used to calculate the change intensity between each target period image and the source period image based on the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image.

[0040] The first processing module is used to determine the initial prediction samples for supervised classification based on the intensity of change and the classification result labels;

[0041] The second processing module is used to purify the initial prediction samples to obtain reliable samples for supervised classification.

[0042] The third processing module is used to perform supervised classification of multiple target period images based on reliable samples to obtain target land cover products.

[0043] In one possible implementation, the second processing module is further configured to:

[0044] Based on the clustering characteristics of the spatial distribution of ground features, broken tags in the initial predicted samples are removed to obtain the first purified sample;

[0045] The first purified sample was characterized to retain the tags whose intensity of change was within the range of change intensity, thus obtaining the second purified sample;

[0046] The Markov random field model and information entropy theory are used to process the second purified sample to obtain a reliable sample for supervised classification.

[0047] In one possible implementation, the second processing module is further configured to:

[0048] The second purified sample was used as a label to pre-classify images of multiple target periods using a support vector machine model to obtain the pre-classification posterior probability and pre-classification results.

[0049] Based on the pre-classification results, a Markov random field model is introduced to output a normalized total energy image;

[0050] For the pre-classification posterior probability, information entropy theory is introduced to calculate the entropy value map;

[0051] Based on the normalized total energy image and entropy map, a weighted map is generated, and a reliable region is determined based on the weighted map;

[0052] The trusted region is spatially overlaid with the second purified sample to obtain a trusted sample for supervised classification.

[0053] In one possible implementation, the first processing module is further configured to:

[0054] Based on the classification results labels and the intensity of change, multiple intensity of change zones were determined;

[0055] For multiple zones with varying intensity, multiple zone thresholds are determined;

[0056] Based on multiple partition thresholds and the intensity of change, the unchanged areas are identified;

[0057] The unchanged regions and classification result labels are overlaid as output to obtain the initial prediction samples for supervised classification.

[0058] In one possible implementation, the second processing module is further configured to:

[0059] Based on the intensity of change and the unchanged area, determine the mean and variance of the intensity of change corresponding to each intensity of change zone;

[0060] Based on the mean and variance, determine the range of change intensity corresponding to each change intensity zone;

[0061] Based on the range of change intensity, the first purified sample is screened to retain the tags whose change intensity is within the range of change intensity, thus obtaining the second purified sample.

[0062] In one possible implementation, the second computing module is further configured to:

[0063] Calculate the spectral slope corresponding to the source period image and the spectral slope corresponding to each target period image respectively;

[0064] Based on the spectral slope of the source period image and the spectral slope of each target period image, calculate the spectral slope vector of the source period image and the spectral slope vector of each target period image.

[0065] Thirdly, this application provides a remote sensing land cover classification device based on transfer learning, comprising: a memory and a processor;

[0066] The memory stores computer-executed instructions;

[0067] When the processor executes the computer execution instructions stored in the memory, the processor performs the first aspect and / or various possible implementations of the first aspect as described above.

[0068] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible embodiments of the first aspect.

[0069] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0070] This application provides a method, apparatus, device, and storage medium for remote sensing land cover classification using transfer learning. The method involves acquiring source period images, corresponding classification result labels for the source period images, and multiple target period images; calculating the spectral slope vectors for the source period images and each target period image; calculating the change intensity between each target period image and the source period image based on these vectors; determining initial prediction samples for supervised classification based on the change intensity and classification result labels; purifying the initial prediction samples to obtain reliable samples for supervised classification; and performing supervised classification on multiple target period images based on these reliable samples to obtain target land cover products. This process begins by acquiring the classification result labels for the source period images and then using these labels... Furthermore, the spectral slope difference method automatically generates and supervises the purification of samples, eliminating the need for extensive manual annotation and significantly reducing the time and manpower costs associated with sample preparation and annotation. Secondly, by utilizing source period images, the classification results of source period images, and multiple target period images, land cover classification products can be quickly generated for multiple target period images by calculating and comparing the intensity of changes in the spectral slope vector. This eliminates the need to recreate samples and classify them for each target period image, significantly improving the temporal resolution of land cover products. Simultaneously, by considering spectral slope differences to detect land cover changes and purifying the initial prediction samples using the spectral slope vector, noise and misclassified samples can be removed, improving sample reliability and thus enhancing classification accuracy. This achieves the effect of improving the temporal resolution, classification efficiency, and accuracy of remote sensing land cover classification products. Attached Figure Description

[0071] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0072] Figure 1 is a schematic diagram of an application data processing system architecture provided in an embodiment of this application;

[0073] Figure 2 is a flowchart illustrating the remote sensing land cover classification method based on transfer learning provided in this embodiment of the application.

[0074] Figure 3 is a schematic flowchart of the remote sensing land cover classification method under transfer learning provided in the embodiments of this application.

[0075] Figure 4 is a flowchart illustrating the remote sensing land cover classification method under transfer learning provided in the embodiments of this application.

[0076] Figure 5 is a schematic diagram of partition threshold segmentation provided in an embodiment of this application;

[0077] Figure 6 is a schematic diagram of the process of remote sensing land cover classification under transfer learning provided in the embodiments of this application.

[0078] Figure 7 is an image diagram showing the results of the remote sensing land cover classification process under transfer learning provided in the embodiments of this application;

[0079] Figure 8 is a schematic diagram of the structure of the remote sensing land cover classification device under transfer learning provided in the embodiment of this application;

[0080] Figure 9 is a schematic diagram of the structure of the remote sensing land cover classification device under transfer learning provided in the embodiment of this application.

[0081] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0082] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0083] Because existing supervised classification methods require manual labeling of different types of land cover samples, which greatly increases time and labor costs, while unsupervised classification methods cannot eliminate interference from phenomena of different objects with the same spectrum and different spectra of the same object, resulting in reduced accuracy, there are technical problems of low efficiency and poor accuracy in remote sensing land classification.

[0084] To address the aforementioned issues, this application provides a remote sensing land cover classification method, apparatus, device, and storage medium based on transfer learning. First, by acquiring classification result labels corresponding to source period images, and automatically generating and supervising purification samples based on these labels and the spectral slope difference method, extensive manual annotation work is eliminated, significantly reducing the time and labor costs associated with sample creation and annotation. Second, by utilizing source period images, their classification result labels, and multiple target period images, land cover classification products can be quickly generated for multiple target period images by calculating and comparing the intensity of changes in the spectral slope vector. This eliminates the need to recreate samples and classify them for each target period image, significantly improving the temporal resolution of the land cover products. Simultaneously, by considering spectral slope differences to detect land cover changes and purifying the initial prediction samples using the spectral slope vector, noise and misclassified samples can be removed, improving sample reliability and thus increasing classification accuracy. This achieves the effect of improving the temporal resolution, classification efficiency, and accuracy of remote sensing land cover classification products.

[0085] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0086] Figure 1 is a schematic diagram of an application data processing system architecture provided in an embodiment of this application. The application data processing system is a computer device. As shown in Figure 1, the architecture includes at least one of a data acquisition device 101, a processing device 102, and a display device 103.

[0087] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the architecture of the application data processing system. In other feasible embodiments of this application, the above architecture may include more or fewer components than illustrated, or combine some components, or split some components, or arrange different components, which can be determined according to the actual application scenario and is not limited here. The components shown in Figure 1 can be implemented in hardware, software, or a combination of software and hardware.

[0088] In the specific implementation process, the data acquisition device 101 may include an input / output interface or a communication interface, and the data acquisition device 101 can be connected to the processing device through the input / output interface or the communication interface.

[0089] The processing device 102 can acquire source period images, classification result labels corresponding to the source period images, and multiple target period images; calculate the spectral slope vector corresponding to the source period images and the spectral slope vector corresponding to each target period image; calculate the change intensity between each target period image and the source period image based on the spectral slope vector corresponding to the source period images and the spectral slope vector corresponding to each target period image; determine the initial prediction samples for supervised classification based on the change intensity and classification result labels; purify the initial prediction samples to obtain reliable samples for supervised classification; and perform supervised classification on multiple target period images based on the reliable samples to obtain target land cover products.

[0090] The display device 103 can also be a touch screen or the screen of a terminal device, used to receive user commands while displaying the above-mentioned content, so as to realize interaction with the user.

[0091] It should be understood that the aforementioned processing device can be implemented by a processor reading instructions from memory and executing those instructions, or it can be implemented by a chip circuit.

[0092] Furthermore, the network architecture and business scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0093] Figure 2 is a flowchart illustrating the remote sensing land cover classification method based on transfer learning provided in this embodiment. As shown in Figure 2, the remote sensing land cover classification method based on transfer learning provided in this embodiment includes:

[0094] S201. Obtain source period images, classification result labels corresponding to source period images, and multiple target period images;

[0095] In this embodiment, the source period image refers to a high-resolution remote sensing image of the reference period used for comparison, and the target period image refers to a high-resolution remote sensing image acquired at the target time point.

[0096] The study area was acquired by obtaining multispectral remote sensing source period images, classification result labels corresponding to the source period images, and multiple target period images.

[0097] S202. Calculate the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image, respectively.

[0098] Spectral analysis was performed on the source period images and the images of each target period to calculate the spectral slope vectors corresponding to the source period images and the images of each target period.

[0099] S203. Calculate the change intensity between each target period image and the source period image based on the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image.

[0100] By comparing the spectral slope vectors corresponding to the source period images obtained in the above steps with the spectral slope vectors corresponding to each target period image, the change intensity between each target period image and the source period image is further calculated. The calculation formula is as follows:

[0101] in, The intensity of the change in the corresponding pixel of the source image and each target image. Let be the spectral slope vector of the source image. S is the spectral slope vector of the image for each target period. (k，k+1) It is the ratio of the difference in spectral reflectance to the difference in the corresponding spectral band.

[0102] S204. Based on the intensity of change and the classification result labels, determine the initial prediction samples for supervised classification;

[0103] The intensity of change in the source image and each target image The zoning threshold method was used to obtain the change intensity thresholds for different land features. Spatial analysis was then used to obtain the unchanged areas and initial samples between the source and target images of the study area.

[0104] Specifically, based on the existing classification result label C1 corresponding to the source image (assuming there are a total of C categories), the intensity of change is... The image is partitioned, and then thresholds are set using the Otsu method (maximum inter-class variance method) for the C intensity variation images after partitioning, to obtain C thresholds t. 1, t2,…,t c After threshold segmentation in the ArcGIS geographic information system platform, changed and unchanged areas are obtained. Using the unchanged areas and the classification result label C1 as input, overlay analysis is performed to obtain the initial sample L required for supervised classification of the target image. i .

[0105] S205. The initial prediction samples are purified to obtain reliable samples for supervised classification.

[0106] Based on the clustering characteristics of spatial distribution of ground features, Markov random field model and information entropy theory, the initial prediction samples are purified to obtain reliable samples that can be used for supervised classification.

[0107] S206. Based on reliable samples, supervised classification is performed on images from multiple target periods to obtain target land cover products.

[0108] Based on reliable samples, the Maximum Likelihood Classification (MLC) method is used to perform supervised classification on images from multiple target periods to obtain the classified target land cover products.

[0109] The remote sensing land cover classification method based on transfer learning provided in this application obtains source period images, corresponding classification result labels for the source period images, and multiple target period images. It calculates the spectral slope vectors for the source period images and each target period image, respectively. Based on these vectors, it calculates the change intensity between each target period image and the source period image. Based on the change intensity and classification result labels, it determines initial prediction samples for supervised classification. The initial prediction samples are then purified to obtain reliable samples for supervised classification. Based on these reliable samples, supervised classification is performed on multiple target period images to obtain target land cover products. This method first obtains the classification result labels for the source period images and then uses these labels and spectral slopes... The differential method automatically generates and supervises the purification of samples, eliminating the need for extensive manual annotation and significantly reducing the time and manpower costs associated with sample creation and annotation. Secondly, by utilizing source period images, their classification labels, and multiple target period images, and calculating and comparing the intensity of changes in spectral slope vectors, land cover classification products can be quickly generated for multiple target period images without requiring the creation and classification of samples for each target period image. This significantly improves the temporal resolution of land cover products. Furthermore, by considering spectral slope differences to detect land cover changes and purifying the initial prediction samples using spectral slope vectors, noise and misclassified samples can be removed, increasing sample reliability and thus improving classification accuracy. This achieves the goal of improving the temporal resolution, classification efficiency, and accuracy of remote sensing land cover classification products.

[0110] Figure 3 is a flowchart illustrating the remote sensing land cover classification method under transfer learning provided in this embodiment. As shown in Figure 3, this embodiment, based on the above embodiments, details the purification process of the initial prediction samples. The method includes:

[0111] S301. Based on the clustering characteristics of the spatial distribution of ground features, remove broken tags from the initial predicted samples to obtain the first purified sample;

[0112] Specifically, considering the clustering characteristics of the spatial distribution of ground features, assuming that each retained sample consists of at least q pixels, raster analysis is used to remove the initial sample L. i The fragmented tags were used to obtain the first purified sample L after preliminary purification. i 1 .

[0113] S302. Based on the intensity of change and the unchanged area, determine the average value and variance of the intensity of change corresponding to each intensity of change partition, and determine the intensity of change range corresponding to each intensity of change partition based on the average value and variance.

[0114] Specifically, the intensity change image obtained using the above-mentioned S203 is... The unchanged areas obtained from S204 above are integrated and analyzed in ArcGIS to obtain the average value of the change intensity of Class C land features. and variance The characteristic statistical method of mean ± N times variance is used to obtain the range of change intensity for each type of land cover.

[0115] S303. Based on the range of change intensity, the first purified sample is screened to retain the tags whose change intensity is within the range of change intensity, thus obtaining the second purified sample;

[0116] Specifically, in ArcGIS, sample L is preserved using the raster calculator. i 1 The pixel intensity variation is included in the label within this range, resulting in the second purified sample L after secondary purification. i 2 .

[0117] S304. Using the second purified sample as a label, the images of multiple target periods are pre-classified using a support vector machine model to obtain the pre-classification posterior probability and pre-classification result.

[0118] Specifically, the sample L after secondary purification i 2 The target image is pre-classified using a Support Vector Machine (SVM) model with labels, and the pre-classification posterior probability P and pre-classification result C are obtained. pre .

[0119] S305. Based on the pre-classification results, a Markov random field model is introduced to output a normalized total energy image.

[0120] Specifically, the Markov random field model refers to the binary Markov random field model or the Gaussian Markov random field model. In the Markov random field model, the probability P of a pixel (m,n) belonging to category X is determined by the spectral information of that pixel and the neighboring pixels. The maximum a posteriori probability problem is transformed into a minimum energy problem by taking the logarithm of the probabilities obtained from the spectral information and the probabilities obtained from the texture information respectively, and then adding them together, as shown in the following formula:

[0121] Where Z is the normalization constant, U C(X(m，n)) and U S(X(m,n)) These are the texture energy and spectral energy of pixel (m,n), respectively. The texture energy is calculated based on the classification information of neighboring pixels. For pixel (m,n), the second-order neighborhood system N(m,n) is defined by the following formula: N(m,n)={(m±1,n),(m,n±1),(m+1,n±1),(m-1,n±1)}

[0122] The texture energy can be obtained through a second-order neighborhood system, and its calculation method is as follows:

[0123] Where β is the weighting coefficient of texture energy, typically ranging from 0.5 to 2, and δ in the above formula... k As shown below:

[0124] Considering that each pixel is independent and that pixels in each class follow a Gaussian distribution, a Gaussian density function is established for each class. Then, for a pixel with a gray value of 's', the probability of it belonging to each class can be obtained. Taking the logarithm of this probability value yields the pixel's spectral energy value, as shown in the following equation:

[0125] Where x is the gray value of a pixel, and μ and σ are the mean and variance of the gray values ​​for each type of land cover. After obtaining the spectral energy value and texture energy value, it is necessary to calculate the spectral energy and texture energy that minimize the total energy. That is, as shown in the following formula: W=arc min(U C(X(m,n)) +U S(X(m,n)) )

[0126] The energy function is minimized using an optimization algorithm based on iterated conditional modes (ICM). The iterative steps are as follows:

[0127] Step 1: Introduce the pre-classification result C from the SVM classifier pre As the initial label for the category;

[0128] Step 2: Traverse the pixels of the image and calculate the spectral energy and texture energy of each pixel when it becomes a different category;

[0129] Step 3: Compare the total energy values ​​of each pixel when it belongs to different categories, and select the state that minimizes the total energy value as the category of that pixel;

[0130] Repeat steps 2 and 3 for a finite number of iterations until convergence, and output the total energy function value at convergence to obtain the maximum posterior probability P.

[0131] S306. For the pre-classification posterior probability, information entropy theory is introduced to calculate the entropy value map;

[0132] Specifically, for the pre-classification posterior probability P, it is first clarified that for a pixel, the larger the difference in probability values ​​between its classifications, the less likely the pixel is to be misclassified; conversely, if the difference in probability values ​​between its classifications is relatively small, the probability of misclassification is greater. By calculating the Shannon entropy of each pixel, it is more difficult to determine the classification of pixels with larger entropy values, while the classification of pixels with smaller entropy values ​​is relatively easier to determine. The definition of Shannon entropy is as follows:

[0133] Among them, (p1,p2,...,p n ) is for the random variable X taking values ​​(x1, x2, ..., x...). n The probability of X taking the values ​​(x1, x2, ..., x) is such that the larger the entropy value, the greater the uncertainty. n The entropy is maximized when the probabilities of ) are equal, as shown below:

[0134] Therefore, the principle of information entropy states that when the situation is unknown, no subjective assumptions are made. In this case, the probability distribution is the most uniform and the uncertainty of prediction is the smallest, and the entropy value is the largest.

[0135] S307. Based on the normalized total energy image and entropy map, generate a weighted map and determine the reliable region based on the weighted map;

[0136] Specifically, for the normalized total energy image I1 and the entropy value image I2, a weighted summation method is used to obtain the weighted image I3, with weights set to 0.6 and 0.4 respectively. Then, the Otsu method is used to automatically segment the probabilistic image I3 to obtain the reliable region. It should be noted that the smaller the pixel value of I3, the easier it is to determine the land cover category and the higher the reliability.

[0137] S308. Spatially overlay the trusted region with the second purified sample to obtain a trusted sample for supervised classification.

[0138] In ArcGIS, the trusted region is compared with the second purified sample L after secondary purification. i2 Spatial overlay analysis was performed to obtain the reliable sample L. i 3 .

[0139] The remote sensing land cover classification method based on transfer learning provided in this application effectively removes noise and irregular small areas by deleting fragmented labels based on the clustering characteristics of land cover spatial distribution, thus obtaining a purer first purified sample. Furthermore, it determines the mean and variance of each change intensity partition based on the change intensity and unchanged areas, thereby determining the change intensity range, which helps to more accurately classify the change characteristics of different areas. Simultaneously, it uses a support vector machine model to pre-classify images from multiple target periods and introduces a Markov random field model and information entropy theory to generate normalized total energy images and entropy maps, respectively. This allows for further analysis of the spatial structure and information distribution of the images, helping to maintain spatial coherence during classification and reduce classification errors. Finally, it determines reliable regions through a weighted map and spatially overlays these regions with the second purified sample to obtain reliable samples for supervised classification, ensuring that the selected samples are more representative. This improves the reliability and processing speed of subsequent classification, achieving the effect of improving the temporal resolution, classification efficiency, and accuracy of remote sensing land cover classification products.

[0140] Figure 4 is a flowchart illustrating the remote sensing land cover classification method under transfer learning provided in this embodiment. As shown in Figure 4, this embodiment, based on the above embodiments, provides a detailed description of the process for obtaining the initial prediction samples, including:

[0141] S401. Based on the classification results labels and change intensity, determine multiple change intensity partitions;

[0142] Specifically, Figure 5 is a schematic diagram of partition threshold segmentation provided in the embodiment of this application. As shown in Figure 5, the intensity of change is partitioned according to the existing classification result labels corresponding to the source image, thereby obtaining multiple intensity of change partitions.

[0143] It should be noted that Figure 5 is only used as an illustrative representation and is not intended as an improvement, nor does it affect the scope of protection of the embodiments of this application.

[0144] S402. For multiple intensity variation zones, determine multiple zone thresholds;

[0145] Specifically, the Otsu method is used to set thresholds for multiple intensity variation images after partitioning, and multiple corresponding thresholds are obtained.

[0146] S403. Based on multiple partition thresholds and change intensity, determine the unchanged areas;

[0147] Specifically, in ArcGIS, threshold segmentation is used to obtain the changed and unchanged areas.

[0148] S404. Overlay the unchanged regions and classification result labels as output to obtain the initial prediction samples for supervised classification.

[0149] Using unchanged regions and classification result labels as input, overlay analysis is performed to obtain the initial samples required for supervised classification of target images.

[0150] The remote sensing land cover classification method provided in this application determines multiple change intensity zones based on classification result labels and change intensity, which can more finely divide areas with different degrees of change in the image. This helps to more accurately analyze and process image data in subsequent steps. In addition, determining the zone threshold for multiple change intensity zones ensures that the change intensity within each zone is within a reasonable range. At the same time, determining the unchanged areas based on the zone thresholds and change intensity can accurately identify the stable and unchanging parts in the image. Finally, the unchanged areas and classification result labels are overlaid to obtain initial prediction samples for supervised classification. These samples contain information on both unchanged areas and classification result labels, further improving their quality and reliability. This improves the processing speed and accuracy of subsequent supervised classification, thereby enhancing the temporal resolution, classification efficiency, and accuracy of remote sensing land cover classification products.

[0151] Figure 6 is a flowchart illustrating the remote sensing land cover classification method under transfer learning provided in this embodiment. As shown in Figure 6, this embodiment, based on the above embodiments, supplements the calculation process of the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image, including:

[0152] S601. Calculate the spectral slope corresponding to the source period image and the spectral slope corresponding to each target period image respectively.

[0153] For source image T1 and image T for each target period i (i = 2, 3, ..., n), calculate the slope between adjacent bands, spectral slope S. (k,k+1) This refers to the ratio of the difference in spectral reflectance to the difference in the corresponding spectral band, and its calculation formula is shown in the following formula:

[0154] Where, ΔR (k,k+1) Δλ represents the difference in reflectivity between adjacent bands. (k,k+1) R represents the wavelength difference between adjacent bands k+1 and k. k+1 and R k These are the reflectances at bands k+1 and k, respectively, and λ k+1 and λ kThis indicates the corresponding wavelength.

[0155] S602. Based on the spectral slope corresponding to the source period image and the spectral slope corresponding to each target period image, calculate the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image.

[0156] For source image T1 and image T for each target period i Combine the spectral slopes into a vector The shape of the spectral curve is described using a slope vector, as shown in the following equation:

[0157] Where m refers to the number of bands in the image, and the above steps yield the source image T1 and the image T for each target period. i The corresponding spectral slope vector and

[0158] The remote sensing land cover classification method based on transfer learning provided in this application extracts and compares spectral features at different time points, providing key spectral information support for monitoring land cover changes, improving classification accuracy, and subsequent land cover identification and analysis tasks. This improves the processing speed and accuracy of subsequent supervised classification, and achieves the effect of improving the temporal resolution, classification efficiency, and accuracy of remote sensing land cover classification products.

[0159] Figure 7 is an image diagram of the results of the remote sensing land cover classification process under transfer learning provided in the embodiments of this application. As shown in Figure 7, the remote sensing land cover classification process provided in the above embodiments is displayed in images step by step through spectral analysis technology, and the corresponding result diagrams can be obtained. Among them, Figure 7a is the source image, Figure 7b is the target image, Figure 7c is the source classification product image, Figure 7d is the change intensity image, Figure 7e is the partition threshold segmentation image, Figure 7f is the first purified sample image, Figure 7g is the second purified sample image, Figure 7h is the SVM pre-classification image, Figure 7i is the entropy image, Figure 7j is the normalized total energy image, Figure 7k is the weighted confidence image, Figure 7l is the confidence sample area image, Figure 7m is the confidence sample image, and Figure 7n is the target classification product image.

[0160] It is understood that Figure 7 is only used as a reference for demonstrating the effect and is not intended as an improvement point, nor does it affect the protection scope of the embodiments of this application.

[0161] Figure 8 is a schematic diagram of the structure of the remote sensing land cover classification device under transfer learning provided in this embodiment of the application. The device in this embodiment can be in the form of software and / or hardware. As shown in Figure 8, the remote sensing land cover classification device 800 under transfer learning provided in this embodiment of the application includes: an acquisition module 801, a first calculation module 802, a second calculation module 803, a first processing module 804, a second processing module 805, and a third processing module 806.

[0162] The acquisition module 801 is used to acquire source period images, classification result labels, and multiple target period images;

[0163] The first calculation module 802 is used to calculate the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image, respectively.

[0164] The second calculation module 803 is used to calculate the change intensity between each target period image and the source period image based on the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image.

[0165] The first processing module 804 is used to determine the initial prediction samples for supervised classification based on the intensity of change and the classification result label;

[0166] The second processing module 805 is used to purify the initial prediction samples to obtain reliable samples for supervised classification.

[0167] The third processing module 806 is used to perform supervised classification of multiple target period images based on reliable samples to obtain target land cover products.

[0168] In one possible implementation, the second processing module 805 is further configured to:

[0169] Based on the clustering characteristics of the spatial distribution of ground features, broken tags in the initial predicted samples are removed to obtain the first purified sample;

[0170] The first purified sample was characterized to retain the tags whose intensity of change was within the range of change intensity, thus obtaining the second purified sample;

[0171] The Markov random field model and information entropy theory are used to process the second purified sample to obtain a reliable sample for supervised classification.

[0172] In one possible implementation, the second processing module 805 is further configured to:

[0173] The second purified sample was used as a label to pre-classify images of multiple target periods using a support vector machine model to obtain the pre-classification posterior probability and pre-classification results.

[0174] Based on the pre-classification results, a Markov random field model is introduced to output a normalized total energy image;

[0175] For the pre-classification posterior probability, information entropy theory is introduced to calculate the entropy value map;

[0176] Based on the normalized total energy image and entropy map, a weighted map is generated, and a reliable region is determined based on the weighted map;

[0177] The trusted region is spatially overlaid with the second purified sample to obtain a trusted sample for supervised classification.

[0178] In one possible implementation, the first processing module 804 is further configured to:

[0179] Based on the classification results labels and the intensity of change, multiple intensity of change zones were determined;

[0180] For multiple zones with varying intensity, multiple zone thresholds are determined;

[0181] Based on multiple partition thresholds and the intensity of change, the unchanged areas are identified;

[0182] The unchanged regions and classification result labels are overlaid as output to obtain the initial prediction samples for supervised classification.

[0183] In one possible implementation, the second processing module 805 is further configured to:

[0184] Based on the intensity of change and the unchanged area, determine the mean and variance of the intensity of change corresponding to each intensity of change zone;

[0185] Based on the mean and variance, determine the range of change intensity corresponding to each change intensity zone;

[0186] Based on the range of change intensity, the first purified sample is screened to retain the tags whose change intensity is within the range of change intensity, thus obtaining the second purified sample.

[0187] In one possible implementation, the second computing module 803 is further configured to:

[0188] Calculate the spectral slope corresponding to the source period image and the spectral slope corresponding to each target period image respectively;

[0189] Based on the spectral slope of the source period image and the spectral slope of each target period image, calculate the spectral slope vector of the source period image and the spectral slope vector of each target period image.

[0190] The remote sensing land cover classification device under transfer learning provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0191] Figure 9 is a schematic diagram of the structure of the remote sensing land cover classification device under transfer learning provided in an embodiment of this application. As shown in Figure 9, the electronic device 900 provided in this embodiment includes at least one processor 901 and a memory 902. Optionally, the device 900 also includes a communication component 903. The processor 901, the memory 902, and the communication component 903 are connected via a bus.

[0192] In a specific implementation, when at least one processor 901 executes computer execution instructions stored in memory 902, at least one processor 901 performs the above-described method.

[0193] The specific implementation process of processor 901 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0194] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0195] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0196] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0197] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0198] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0199] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0200] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0201] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0202] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0203] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0204] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0205] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0206] Finally, it should be noted that other embodiments of this application will readily conceive of by those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A remote sensing land cover classification method based on transfer learning, characterized in that, include: Acquire source period images, classification result labels corresponding to the source period images, and multiple target period images; Calculate the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image, respectively; Based on the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image, calculate the change intensity between each target period image and the source period image; Based on the classification result labels corresponding to the source period images, the change intensity is divided into regions. Then, multiple thresholds are set using the maximum inter-class variance method for the multiple change intensity images after partitioning. After segmentation using multiple thresholds, changed and unchanged regions are obtained. The unchanged regions and the classification result labels are used as input for overlay analysis to obtain the initial prediction samples required for supervised classification of the target image. The initial prediction samples are then subjected to secondary purification and confidence assessment to obtain reliable samples for supervised classification. Based on the reliable samples, supervised classification is performed on the multiple target period images to obtain the target land cover product; The process of performing secondary purification and confidence assessment on the initial predicted samples to obtain reliable samples for supervised classification includes: Based on the clustering characteristics of the spatial distribution of ground features, broken tags in the initial predicted samples are removed to obtain the first purified sample; The first purified sample is subjected to feature analysis to retain the tags whose change intensity is within the range of change intensity, thus obtaining the second purified sample; The credibility of the second purified sample is determined by using the Markov random field model and information entropy theory to obtain a credible sample for supervised classification.

2. The method according to claim 1, characterized in that, The process employs a Markov random field model and information entropy theory to determine the credibility of the second purified sample, thereby obtaining a credible sample for supervised classification, including: The second purified sample is used as a label to perform pre-classification processing on the multiple target period images using a support vector machine model to obtain the pre-classification posterior probability and pre-classification result; Based on the pre-classification results, a Markov random field model is introduced to output a normalized total energy image; For the pre-classification posterior probability, information entropy theory is introduced to calculate the entropy value map; Based on the normalized total energy image and the entropy map, a weighted map is generated, and a reliable region is determined based on the weighted map; The trusted region is spatially superimposed with the second purified sample to obtain a trusted sample for supervised classification.

3. The method according to claim 1, characterized in that, The process involves partitioning the change intensity based on the classification result labels corresponding to the source period image, then applying the maximum inter-class variance method to set multiple thresholds for the partitioned change intensity images, segmenting them using multiple thresholds to obtain changed and unchanged regions, and using the unchanged regions and the classification result labels as input for overlay analysis to obtain the initial prediction samples required for supervised classification of the target image, including: Based on the classification result labels and the intensity of change, multiple intensity of change zones are determined; For the multiple intensity variation zones, multiple zone thresholds are determined; Based on the multiple partition thresholds and the intensity of change, the unchanged regions are determined; The unchanged regions and the classification result labels are overlaid as output to obtain initial prediction samples for supervised classification.

4. The method according to claim 3, characterized in that, The step of performing feature analysis on the first purified sample to retain tags with change in intensity within the range of change in intensity, to obtain the second purified sample, includes: Based on the intensity of change and the unchanged region, determine the average and variance of the intensity of change corresponding to each intensity of change partition; Based on the average value and the variance, determine the range of change intensity corresponding to each of the change intensity zones; Based on the range of change intensity, the first purified sample is screened to retain tags whose change intensity falls within the range of change intensity, thus obtaining a second purified sample.

5. The method according to any one of claims 1 to 4, characterized in that, The calculation of the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image includes: Calculate the spectral slope corresponding to the source period image and the spectral slope corresponding to each target period image, respectively; Based on the spectral slope corresponding to the source period image and the spectral slope corresponding to each target period image, calculate the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image.

6. A remote sensing land cover classification device based on transfer learning, characterized in that, include: The acquisition module is used to acquire source period images, the classification result labels, and multiple target period images; The first calculation module is used to calculate the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image, respectively. The second calculation module is used to calculate the change intensity between each target period image and the source period image based on the spectral slope vector corresponding to the source period image and the spectral slope vector corresponding to each target period image. The first processing module is used to partition the change intensity according to the classification result label corresponding to the source period image, and then use the maximum inter-class variance method to set multiple thresholds for the partitioned multiple change intensity images. After segmentation using multiple thresholds, the changed and unchanged regions are obtained. The unchanged regions and the classification result label are used as input for overlay analysis to obtain the initial prediction samples required for supervised classification of the target image. The second processing module is used to perform two purification processes and a confidence judgment on the initial prediction sample to obtain a reliable sample for supervised classification. The third processing module is used to perform supervised classification on the multiple target period images based on the trusted samples to obtain the target land cover product. The second processing module is specifically used to delete broken labels in the initial prediction sample based on the clustering characteristics of the spatial distribution of ground features to obtain a first purified sample; to perform feature analysis on the first purified sample to retain labels whose change intensity is within the range of change intensity to obtain a second purified sample; and to use a Markov random field model and information entropy theory to determine the credibility of the second purified sample to obtain a credible sample for supervised classification.

7. A remote sensing land cover classification device based on transfer learning, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform the remote sensing land cover classification method under transfer learning as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the remote sensing land cover classification method under transfer learning as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the remote sensing land cover classification method under transfer learning as described in any one of claims 1 to 5.

Images