Remote sensing cross-scene mobility measurement method and related device
By constructing feature space, extracting weighted feature differences in remote sensing images in semantic segmentation task, performing comparison and analysis, the problems of multi-category feature confusion and dimensional feature contributions in remote sensing images are solved, and the accuracy and accuracy of model migration are improved.
Patent Information
- Application Number
- CN202411955566.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-28
- Publication Date
- 2025-04-25
AI Technical Summary
The existing migration ability metrics have problems such as confusion of multiple categories of features and unequal contributions to different dimensions of feature in the semantic segmentation task of remote sensing images, resulting in inaccurate model migration.
By constructing the feature space of source domain and target domain, extracting the foreground and background types, calculating weighted feature differences, and applying weights according to the importance of different dimension features, calculating the migration metrics and accuracy losses, and performing comparative analysis.
It effectively avoids confusion of category features, improves the accuracy of feature differences, enhances the accuracy prediction ability after model migration, and improves the stable correlation between the migration measurement and the accuracy loss after migration.
Smart Images
Figure CN119888515A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of remote sensing technology, and in particular relates to a remote sensing cross-scene transferability measurement method and related devices. Background Art
[0002] Satellite remote sensing observation of the earth is periodic, so a large amount of historical image data has been accumulated in various scenarios. Therefore, the migration problem of artificial intelligence models in the field of satellite remote sensing belongs to the category of domain adaptation, that is, the algorithm model can obtain the data features of the target domain during the modeling process. Transferability measurement is one of the core steps in modeling domain adaptive models. Current research in this field focuses on the description of the feature space of the source domain and the target domain and the measurement of spatial similarity.
[0003] In today's technical field, transferability measurement methods have important research significance and application value in many aspects. In particular, in the research direction of transferability measurement methods based on the difference in data features between the source domain and the target domain, it has promoted the development of related technologies to a certain extent. This method analyzes and derives the transferability of the neural network model from the data feature level. Since it avoids a large number of model retraining processes, its computational efficiency is relatively high, which is very beneficial in many scenarios.
[0004] However, this method has exposed many shortcomings in the application scenario of remote sensing image semantic segmentation tasks. First of all, the current transferability metric research almost all focuses on image classification tasks, and lacks sufficient consideration for remote sensing image semantic segmentation tasks. The remote sensing field has its own unique complexity and particularity, which is significantly different from ordinary image classification tasks. For example, remote sensing images often cover large geographical areas, contain a wide variety of ground features, and the representation of different ground features in the images is complex and varied. Due to the lack of application scenario modeling for remote sensing image semantic segmentation tasks, these methods cannot be directly applied to model migration in the remote sensing field, which greatly limits the further development of remote sensing technology in related fields.
[0005] Secondly, there is a problem of confusing features of multiple categories when constructing feature space. In many existing methods, the average of features of multiple categories is taken. Although this method may be reasonable in some scenarios, it is very inappropriate in the field of remote sensing ground feature classification, especially when the features of foreground and background classes are significantly different. Among remote sensing ground features, the features of foreground classes (such as buildings, vegetation, etc.) and background classes (such as roads, water bodies, etc.) in images are often very different. Fusion of their features will result in the inability to accurately distinguish the influence of different categories when performing transferability measurement, which in turn affects the accuracy and effectiveness of model migration.
[0006] Furthermore, existing methods ignore the importance of different dimensions in the feature space. When calculating the overall domain difference, the contribution of all dimensional differences is considered equal, which is an oversimplified approach. In remote sensing images, features of different dimensions have different importance for the classification and identification of land features. For example, features of certain dimensions may be more critical for distinguishing vegetation types, while other dimensions may be important in identifying water bodies. Ignoring this difference will cause the final transferability metric to fail to capture the differences in key dimensions, and thus fail to produce a stable correlation with precision loss, making it difficult to accurately predict and control changes in precision during model migration. Summary of the invention
[0007] The main purpose of the present invention is to provide a remote sensing cross-scene transferability measurement method and related devices, aiming to solve the problem of different dimensions in the existing confusing features of multiple categories such as foreground and background classes and feature space.
[0008] To achieve the above object, the present invention provides a remote sensing cross-scene transferability measurement method, comprising:
[0009] Construct source domain and target domain feature space;
[0010] Extracting foreground features of the source domain and the target domain from the feature spaces of the source domain and the target domain, and calculating the feature difference between the source domain and the target domain;
[0011] According to the segmentation parameters of the segmentation head in the source domain semantic segmentation model, weights are applied to the feature differences of the corresponding feature dimensions to obtain the weighted feature differences between the source domain and the target domain;
[0012] The transferability metric is calculated based on the weighted feature difference, and the accuracy loss after the source domain model is migrated to the target domain is compared and analyzed.
[0013] Furthermore, the step of “constructing source domain and target domain feature space” specifically includes:
[0014] According to the remote sensing cross-scene migration task, collating the source domain and the target domain datasets;
[0015] Training a source domain binary classification semantic segmentation model according to the source domain dataset;
[0016] Inputting the images in the source domain and the target domain datasets into the source domain binary classification semantic segmentation model;
[0017] The source domain and the target domain feature space are obtained.
[0018] Furthermore, the step of “calculating the feature difference between the source domain and the target domain” specifically includes:
[0019] Extracting the source domain and target domain foreground class feature space according to the pixel-by-pixel labels of the source domain and target domain datasets;
[0020] Calculate the feature mean of each dimension of the source domain foreground class feature space;
[0021] Calculate the feature space mean of the foreground class in each batch of the target domain according to the predetermined batches;
[0022] The mean of the features of each batch in the target domain is subtracted from the mean of the features of each dimension in the source domain to obtain the relative feature difference.
[0023] As a solution of the present invention, the step of "applying weights to feature differences of corresponding feature dimensions" specifically includes:
[0024] Extracting target parameters of a segmentation head in the source domain binary classification semantic segmentation model;
[0025] The target parameters are used to apply weights to the feature differences of each dimension between the source domain and the target domain, respectively, to obtain weighted feature differences of each batch of the source domain and the target domain.
[0026] Specifically, the target parameters at least include weight parameters and bias parameters of the segmentation head in the source domain binary classification semantic segmentation model.
[0027] As another solution of the present invention, the step of "calculating the transferability metric index and the accuracy loss after the source domain model is migrated to the target domain, and performing transferability metric comparative analysis" specifically includes:
[0028] Calculating a transferability metric for each batch from the source domain to the target domain according to the weighted feature difference;
[0029] Inputting the source domain data set into the source domain binary classification semantic segmentation model to obtain the source domain accuracy of the entire source domain data set;
[0030] The target domain dataset is input into the source domain binary classification semantic segmentation model in batches to obtain the target domain accuracy of each batch.
[0031] Subtract the source domain accuracy from the target domain accuracy to obtain the accuracy loss after migration;
[0032] The transferability metric index and the precision loss are combined to perform a transferability metric comparison analysis.
[0033] As another solution of the present invention, it is characterized in that the step of "comparative analysis of transferability metrics" specifically includes:
[0034] When the transferability metric and the precision loss are greater than a preset threshold, the source domain binary classification semantic segmentation model has no transfer value;
[0035] When the transferability metric is greater than a preset threshold, but the accuracy loss is less than a preset threshold, the training sample size of the source domain binary classification semantic segmentation model is increased and overfitting is reduced;
[0036] When the transferability metric and the precision loss are both less than a preset threshold, the source domain binary classification semantic segmentation model can be migrated.
[0037] A remote sensing cross-scene transferability measurement device includes: a processing unit, configured to call a target remote sensing cross-scene transferability measurement method to perform remote sensing cross-scene transferability measurement on a source domain binary classification semantic segmentation model; wherein the remote sensing cross-scene transferability measurement method is any of the remote sensing cross-scene transferability measurement methods described above.
[0038] A computer-readable storage medium, wherein the computer instructions are used to cause the computer to execute any one of the remote sensing cross-scene transferability measurement methods described above.
[0039] An electronic device comprises: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute any one of the remote sensing cross-scene transferability measurement methods described above.
[0040] In the above technical solution of the present invention, by constructing application scenarios according to remote sensing cross-scene migration tasks, the uniqueness of remote sensing images is fully taken into account; and the foreground and background classes of remote sensing objects are further distinguished in the feature space, avoiding the influence of different category feature confusion on the effectiveness of the transferability metric; at the same time, weights are applied to the feature differences of each dimension in the feature space, and reasonable adjustments are made according to the importance of different dimensional features in the classification of object elements, so that the dimensions that contribute greatly to the classification results also have a great impact on the transferability metric, thereby effectively improving the stable correlation between the transferability metric and the accuracy loss after migration. The present invention is a new neural network cross-scene transferability metric method for intelligent classification of surface elements, which solves the above-mentioned key problems that hinder the development of transferability metrics in the field of remote sensing, thereby guiding the research and development and application of migration models in larger areas and more scales of remote sensing, and providing strong technical support for the application of remote sensing technology in a wider range of fields. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.
[0042] Figure 1 A schematic diagram of the overall steps of a remote sensing cross-scene transferability measurement method according to an embodiment of the present invention;
[0043] Figure 2 A schematic diagram of the steps of constructing a feature space according to another embodiment of the present invention;
[0044] Figure 3 A schematic diagram of the steps for calculating feature differences according to an embodiment of the present invention;
[0045] Figure 4 Schematic diagram of feature difference weighting steps in an embodiment of the present invention;
[0046] Figure 5 Schematic diagram of the steps of calculating and analyzing the portability metric according to an embodiment of the present invention. DETAILED DESCRIPTION
[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0048] It should be noted that all directional indications (such as up, down, etc.) in the embodiments of the present invention are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.
[0049] Furthermore, the technical solutions between the various embodiments of the present invention may be combined with each other, but this must be based on the fact that they can be implemented by ordinary technicians in the field. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.
[0050] See also Figure 1 According to one aspect of the present invention, a remote sensing cross-scene transferability measurement method is provided, comprising:
[0051] S1. Construct feature space, that is, construct source domain and target domain feature space;
[0052] like Figure 2 As shown, in this step, the existing cross-scenario migration task of the neural network model for remote sensing surface element classification is specifically included, the source domain and the target domain data sets are sorted, and the binary classification semantic segmentation data sets of the source domain (region A) and the target domain (region B) are sorted out, including images and labels, taking the cross-region migration scenario as an example. ; The source domain binary classification semantic segmentation model is trained according to the source domain data set, and the binary classification semantic segmentation neural network model of the source domain is trained by taking the U-Net network as an example; the images in the source domain and the target domain data sets are input into the source domain binary classification semantic segmentation model to obtain the source domain and the target domain feature space; according to the existing research practice, the features of the second to last layer of the model are taken, that is, the features of the input segmentation head, so as to obtain the feature space of the source domain and the target domain.
[0053] S2. Extracting foreground features of the source domain and the target domain from the feature space of the source domain and the target domain, and calculating the feature difference between the source domain and the target domain. By separating the features of the foreground class and the background class in the feature space, the influence of the confusion of features of different categories on the effectiveness of the transferability metric is avoided, and the prediction level of the accuracy after the model transfer is effectively improved.
[0054] Specific as Figure 3 As shown, this step contains three sub-steps
[0055] Step 1: Extract the foreground feature space of the source domain and the target domain according to the pixel-by-pixel labels of the source domain and the target domain datasets; specifically, according to the pixel-by-pixel labels of the source domain and the target domain datasets, discard the background features in the feature space, extract and sort the foreground features, and obtain the foreground feature space of the source domain and the target domain, which are respectively denoted as Z s , Z t (s represents the source domain, i.e. source domain; t represents the target domain, i.e. target domain).
[0056]
[0057] N is the total number of pixels of the foreground class in the source domain, and 64 is the number of feature dimensions;
[0058]
[0059] M is the total number of pixels of the foreground class in the target domain.
[0060] Step 2: Calculate the mean of each dimension of the source domain foreground feature space; specifically, for the source domain foreground feature space Z sIn multiple feature dimensions (this embodiment takes 64 feature dimensions as an example, but it does not mean that the technical solution of the present invention can only implement 64 feature dimensions), calculate the average value of all features in each dimension, denoted as μ s , the calculation formula is as follows:
[0061]
[0062] The target domain is calculated in predetermined batches according to the feature mean of the foreground feature space of each batch; for the foreground feature space Z of the target domain t , according to the foreground class features of each batch containing a predetermined number of images, they are divided into P batches (this embodiment takes each batch containing 4 images as an example, but it does not mean that the technical solution of the present invention can only be implemented with 4 images per batch), and each batch is regarded as a sub-target domain for migration. For the kth (1≤k≤P) batch, the total number of foreground pixels in the batch is recorded as m k , calculate the average value of all features of the 64 feature dimensions in the foreground feature space, denoted as μ t,k ; At the same time, the mean of all batches in the target domain is μ t , the calculation formula is as follows:
[0063]
[0064] Step 3: Subtract the feature mean of each batch in the target domain from the feature mean of each dimension in the source domain to obtain the relative feature difference; specifically, the feature mean μ of the kth (1≤k≤P) batch in the target domain is t,k and the feature mean μ of the entire source domain s The relative characteristic difference is recorded as Δμ k , while keeping the relative characteristic difference of all batches as Δμ, retaining 64 feature dimensions, the calculation formula is as follows:
[0065]
[0066] S3. According to the segmentation parameters of the segmentation head in the source domain semantic segmentation model, weights are applied to the feature differences of the corresponding feature dimensions to obtain the weighted feature differences between the source domain and the target domain. By applying weights to the multi-dimensional feature space differences between the source domain and the target domain, the dimensions that contribute greatly to the classification results also have a greater impact on the transferability metric, fully considering the importance of features of different dimensions, and effectively improving the stable correlation between the transferability metric and the accuracy loss after migration.
[0067] Specific as Figure 4 As shown, this step contains two sub-steps
[0068] Step 1: Extract the target parameters of the segmentation head in the source domain binary classification semantic segmentation model; wherein the target parameters at least include the weight parameters and bias parameters of the segmentation head in the source domain binary classification semantic segmentation model. Specifically, based on the trained source domain binary classification semantic segmentation model, extract the weight parameters and bias parameters of the model segmentation head, denoted as W and B respectively, as follows:
[0069]
[0070] Among them, c represents the category, and in the binary classification there are only foreground and background categories; d represents the dimension, that is, 64 feature dimensions.
[0071] Step 2: Use the target parameters to apply weights to the feature differences of each dimension between the source domain and the target domain, and obtain the weighted feature differences of each batch of the source domain and the target domain. Specifically, based on the 64-dimensional weights and biases of the source domain model segmentation head, weights are applied to the 64-dimensional feature differences of the source domain and the target domain, so as to obtain the weighted feature differences of all batches of the source domain and the target domain, which are denoted as G. The details are as follows:
[0072]
[0073] g k,d =Δμ k,d ×[(w 0,d -w 1,d )×10+(b 0 -b 1 )],Δμ k,d ∈Δμ;
[0074] Among them, g k,d represents the weighted feature difference of the dth dimension of the kth batch, Δμ k,d Represents the relative feature difference of the dth dimension of the kth batch in core step 2.
[0075] S4. Calculate the transferability metric based on the weighted feature difference, and perform a comparative analysis on the accuracy loss after the source domain model is migrated to the target domain.
[0076] Specific as Figure 5 As shown, this step contains three sub-steps
[0077] Step 1: Calculate the transferability metric index of each batch from the source domain to the target domain according to the weighted feature difference; Specifically, based on the weighted feature difference G obtained in the above step, calculate the transferability metric index of multiple batches from the source domain to the target domain, denoted as S, as follows:
[0078]
[0079] Step 2: Input the source domain dataset into the source domain binary classification semantic segmentation model to obtain the source domain accuracy of the entire source domain dataset; specifically, input the source domain dataset into the trained source domain binary classification semantic segmentation model to obtain the accuracy of the entire source domain dataset, denoted as F s According to the batches divided when calculating the average value of the target domain features, the target domain dataset is input into the trained source domain binary classification semantic segmentation model in batches, and the accuracy of the target domain for a total of P batches is obtained, denoted as F t .in,
[0080]
[0081] The difference between the source domain accuracy and the target domain accuracy is used to obtain the accuracy loss after migration, which is denoted as ΔF. The calculation formula is as follows:
[0082]
[0083] Δf k =F s -f t,k ,f t,k ∈F t ;
[0084] Step 3: Combine the transferability metric and the precision loss to perform a transferability metric comparison analysis. and By conducting a joint comparative analysis, we can determine the model's transferability and provide guidance and suggestions for the next step of migration. Specifically, it can be divided into the following situations:
[0085] When the transferability metric and the precision loss are greater than a preset threshold, the source domain binary classification semantic segmentation model has no transfer value; that is, if S and ΔF are both relatively large, it means that the source domain and the target domain have large feature differences and the precision loss is also large, so the source domain model has no transfer value;
[0086] When the transferability metric is greater than a preset threshold, but the precision loss is less than a preset threshold, the training sample size of the source domain binary classification semantic segmentation model is increased and overfitting is reduced; that is, if S is large but ΔF is small, it means that the precision loss is large when the feature difference between the source domain and the target domain is small. It is speculated that the source domain model is overfitted in the source domain and needs to be adjusted before migration. It is recommended to increase the sample size and reduce overfitting when training the model in the source domain.
[0087] When the transferability metric and the precision loss are both less than a preset threshold, the source domain binary classification semantic segmentation model can be migrated; that is, if S and ΔF are both small, it means that the feature difference between the source domain and the target domain is small and the precision loss is small, and the model can be directly migrated.
[0088] The present invention also includes a remote sensing cross-scene transferability measurement device, including: a processing unit, configured to call a target remote sensing cross-scene transferability measurement method to perform remote sensing cross-scene transferability measurement on a source domain binary classification semantic segmentation model; wherein the remote sensing cross-scene transferability measurement method is any one of the remote sensing cross-scene transferability measurement methods described above.
[0089] As an improved solution of the present invention, it also includes a computer-readable storage medium, and the computer instructions are used to enable the computer to execute any of the remote sensing cross-scene transferability measurement methods described above.
[0090] As another improved solution of the present invention, it also includes an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute any one of the remote sensing cross-scene transferability measurement methods described above.
[0091] In the above technical solution of the present invention, by constructing application scenarios according to remote sensing cross-scene migration tasks, the uniqueness of remote sensing images is fully taken into account; and the foreground and background classes of remote sensing objects are further distinguished in the feature space, avoiding the influence of different category feature confusion on the effectiveness of the transferability metric; at the same time, weights are applied to the feature differences of each dimension in the feature space, and reasonable adjustments are made according to the importance of different dimensional features in the classification of object elements, so that the dimensions that contribute greatly to the classification results also have a great impact on the transferability metric, thereby effectively improving the stable correlation between the transferability metric and the accuracy loss after migration. The present invention is a new neural network cross-scene transferability metric method for intelligent classification of surface elements, which solves the above-mentioned key problems that hinder the development of transferability metrics in the field of remote sensing, thereby guiding the research and development and application of migration models in larger areas and more scales of remote sensing, and providing strong technical support for the application of remote sensing technology in a wider range of fields.
[0092] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. All equivalent structural changes made using the contents of the present invention's specification and drawings, or directly / indirectly applied in other related technical fields, are included in the patent protection scope of the present invention.
Claims
1. A remote sensing cross-scene transferability measurement method, characterized in that: include Construct source domain and target domain feature space; Extracting foreground features of the source domain and the target domain from the feature spaces of the source domain and the target domain, and calculating the feature difference between the source domain and the target domain; According to the segmentation parameters of the segmentation head in the source domain semantic segmentation model, weights are applied to the feature differences of the corresponding feature dimensions to obtain the weighted feature differences between the source domain and the target domain; The transferability metric is calculated based on the weighted feature difference, and the accuracy loss after the source domain model is migrated to the target domain is compared and analyzed.
2. The remote sensing cross-scene transferability measurement method according to claim 1, characterized in that: The step of "constructing source domain and target domain feature space" specifically includes: According to the remote sensing cross-scene migration task, collating the source domain and the target domain datasets; Training a source domain binary classification semantic segmentation model according to the source domain dataset; Inputting the images in the source domain and the target domain datasets into the source domain binary classification semantic segmentation model; The source domain and the target domain feature space are obtained.
3. The remote sensing cross-scene transferability measurement method according to claim 1, characterized in that: The step of "calculating the feature difference between the source domain and the target domain" specifically includes: Extracting the source domain and target domain foreground class feature space according to the pixel-by-pixel labels of the source domain and target domain datasets; Calculate the feature mean of each dimension of the source domain foreground class feature space; Calculate the feature space mean of the foreground class in each batch of the target domain according to the predetermined batches; The mean of the features of each batch in the target domain is subtracted from the mean of the features of each dimension in the source domain to obtain the relative feature difference.
4. The remote sensing cross-scene transferability measurement method according to claim 1, characterized in that: The step of "applying weights to the feature differences of the corresponding feature dimensions" specifically includes: Extracting target parameters of a segmentation head in the source domain binary classification semantic segmentation model; The target parameters are used to apply weights to the feature differences of each dimension between the source domain and the target domain, respectively, to obtain weighted feature differences of each batch of the source domain and the target domain.
5. The remote sensing cross-scene transferability measurement method according to claim 4, characterized in that: The target parameters at least include weight parameters and bias parameters of the segmentation head in the source domain binary classification semantic segmentation model.
6. The remote sensing cross-scene transferability measurement method according to claim 5, characterized in that: The step of "calculating the transferability metric index and the accuracy loss after the source domain model is migrated to the target domain, and performing a comparative analysis of the transferability metric" specifically includes: Calculating a transferability metric for each batch from the source domain to the target domain according to the weighted feature difference; Inputting the source domain data set into the source domain binary classification semantic segmentation model to obtain the source domain accuracy of the entire source domain data set; The target domain dataset is input into the source domain binary classification semantic segmentation model in batches to obtain the target domain accuracy of each batch. Subtract the source domain accuracy from the target domain accuracy to obtain the accuracy loss after migration; The transferability metric index and the precision loss are combined to perform a transferability metric comparison analysis.
7. The remote sensing cross-scene transferability measurement method according to claim 6, characterized in that: The step of "comparative analysis of transferability metrics" specifically includes: When the transferability metric and the precision loss are greater than a preset threshold, the source domain binary classification semantic segmentation model has no transfer value; When the transferability metric is greater than a preset threshold, but the accuracy loss is less than a preset threshold, the training sample size of the source domain binary classification semantic segmentation model is increased and overfitting is reduced; When the transferability metric and the precision loss are both less than a preset threshold, the source domain binary classification semantic segmentation model can be migrated.
8. A remote sensing cross-scene transferability measurement device, characterized in that: include: The processing unit is configured to call the target remote sensing cross-scene transferability measurement method to perform remote sensing cross-scene transferability measurement on the source domain binary classification semantic segmentation model; wherein the remote sensing cross-scene transferability measurement method is the remote sensing cross-scene transferability measurement method described in any one of claims 1-7.
9. A computer-readable storage medium, characterized in that: The computer instructions are used to enable the computer to execute the remote sensing cross-scene transferability measurement method described in any one of claims 1-7.
10. An electronic device, characterized in that: include: at least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the remote sensing cross-scene transferability measurement method described in any one of claims 1-7.