ConvNeXt-OrthoKAN model-based satellite remote sensing image classification method

Abstract

The invention provides a satellite remote sensing image classification method based on a ConvNext-OrthoKAN model. According to the method, the design of a pre-training model and a novel classifier is combined. According to the core technical scheme, firstly, multi-scale features are extracted through a pre-trained ConvNeXt-Tiny network, and the general feature representation capability is reserved by fixing parameters of a convolutional layer of the first half part of the multi-scale features; secondly, a dynamic classifier is designed, and Hermite, Laguerre and Jacobi orthogonal polynomial functions are adopted as KAN primary functions; the classifier can flexibly switch primary function configuration according to different topographic features. In the training process, a two-stage strategy is adopted: firstly, parameters of a feature extraction network are solidified so as to reduce resource consumption; and then independent parameter optimization is carried out on the classifier module. According to the method, while the model training complexity is reduced, the feature discrimination capability and scene adaptability of the multi-category satellite remote sensing images are effectively improved.

Description

Technical Field

[0001] This invention relates to remote sensing image processing, belonging to the field of computer vision. Specifically, it relates to a remote sensing image classification method that integrates frozen pre-trained convolutional features (ConvNeXt-Tiny) with an orthogonal polynomial basis network (Orthogonal-KAN). This invention constructs a two-stage modeling framework by combining ConvNeXt's transfer feature optimization with the basis function combination mechanism of OrthoKAN. The framework is suitable for target recognition and classification in satellite remote sensing images under complex backgrounds. Background Art

[0002] In recent years, the rapid development of satellite technology has driven the widespread use of high-resolution remote sensing imagery. The integration of multi-satellite time series data has significantly improved the accuracy of specific event detection and long-term trend analysis, providing important support for global climate model research and ecological and environmental monitoring. Satellite systems such as Landsat, Sentinel, and MODIS have been widely used for land cover classification, vegetation monitoring, and atmospheric environment tracking. The introduction of deep learning technology has further accelerated breakthroughs in the performance of remote sensing imagery in tasks such as object detection and classification.

[0003] However, efficient processing of high-resolution remote sensing imagery still faces severe challenges. Existing deep neural networks face problems such as high training costs and surging memory requirements due to their large number of parameters. Although some model compression techniques (such as pruning and quantization) can partially alleviate the above contradictions, they often come at the expense of model accuracy, and the ability to fuse multi-scale features in complex scenarios is still insufficient. For example, when processing high-resolution images, convolutional neural networks and visual transformers rely on fixed-level feature extraction mechanisms, making it difficult to balance the association between local details and global semantics, resulting in reduced accuracy in small-scale target recognition and loss of edge information. In addition, the overfitting phenomenon caused by over-parameterization of the model further limits its generalization performance in dynamic environments.

[0004] Existing traditional remote sensing image classification methods usually use CNN or Transformer architectures. However, under complex background interference, they have the following shortcomings:

[0005] (1) The fully connected layer has limited feature expression capabilities and cannot effectively capture high-order nonlinear relationships in remote sensing data;

[0006] (2) The fixed activation function pattern leads to poor adaptability to small-scale targets;

[0007] Conventional classification methods are unable to effectively address the multiple challenges of target morphological diversity, scale differences, and background noise in satellite remote sensing images. Therefore, there is an urgent need to develop a remote sensing image classification method based on a novel network architecture that can improve classification accuracy and model robustness while reducing computational resource consumption. Summary of the Invention

[0008] Technical problems to be solved

[0009] To address the problems of diverse target morphologies, sensitivity to noise interference, and poor adaptability to small-scale targets in satellite remote sensing image classification, the present invention provides a satellite remote sensing image classification method based on the ConvNeXt-OrthoKAN model. This method constructs an efficient remote sensing image classification model by inputting features extracted by ConvNeXt-Tiny into Orthogonal-KAN (a KAN network structure that uses Hermite, Laguerre, and Jacobi orthogonal polynomials as basis functions). This significantly improves the classification accuracy and model generalization ability of multi-category satellite scene data.

[0010] Technical Solution

[0011] In order to achieve the above-mentioned purpose, the present invention generally adopts the following technical solutions:

[0012] S1: Construct the ConvNext-OrthoKAN model. This model consists of a feature extraction module (ConvNext-Tiny) and a multi-basis function classifier module (Orthogonal-KAN). Orthogonal-KAN is a KAN network structure that uses Hermite, Laguerre, and Jacobi orthogonal polynomials as basis functions.

[0013] S2: Model training and generation. The ConvNext-OrthoKAN model is trained using the remote sensing image dataset to obtain the trained ConvNext-OrthoKAN model.

[0014] S3: Generate classification results of the remote sensing image to be processed. Use the trained ConvNext-OrthoKAN model to complete the processing of the remote sensing image to be processed and obtain the classification results of the remote sensing image to be processed.

[0015] S1 Construction of Orthogonal-KAN Model

[0016] The model consists of a feature extraction module and a multi-basis function classifier module, specifically including:

[0017] S11 feature extraction module

[0018] The pre-trained ConvNeXt-Tiny architecture is used as the feature extraction backbone network to extract multi-scale satellite remote sensing image features and freeze the pre-trained parameters to reduce computational overhead. The mathematical representation of feature extraction is:

[0019] (1)

[0020] In the formula is the feature extraction function (where subscript Represents Extraction), ConvBlock is depthwise separable convolution stages, the final output dimension is The feature tensor of 、 Indicates the height and width of the image, is the number of output channels of the last convolutional layer of ConvNeXt-Tiny).

[0021] S12 multi-basis function classifier module

[0022] A configurable KAN architecture was designed, using a KAN layer composed of orthogonal polynomial basis functions for nonlinear transformations. Three independent classifier modules were developed, each integrating a different orthogonal polynomial basis function. The polynomials of the different basis functions were subjected to Tanh input normalization. The multi-basis function classifier module uses Hermite / Laguerre / Jacobi orthogonal polynomials as the KAN basis functions, and the classifier output is:

[0023] (2)

[0024] In the formula is the highest order, is an orthogonal polynomial basis of set degree (Hermite, Laguerre or Jacobi), The KAN subscript poly indicates that the nonlinear basis function type of the classifier is an orthogonal polynomial, which is used to distinguish KAN classifiers with different basis function types. The step-by-step recursive relationship of different basis functions is as follows:

[0025] Hermite basis functions:

[0026] (3)

[0027] Laguerre polynomials (with parameters ):

[0028] (4)

[0029] Jacobi polynomials (with parameters ):

[0030] (5)

[0031] in, is a parameter that controls the weight distribution of the polynomial. By adjusting The value of can change the local sensitive area of ​​the basis function on the input feature normalization area, making it more adaptable to the statistical characteristics of the remote sensing image features.

[0032] S2 model training and generation

[0033] S21 Initialize ConvNeXt-OrthoKAN

[0034] Using a pre-trained ConvNeXt-Tiny backbone network, we freeze the parameters of the first three stages to maintain the stability of the pre-trained parameters, and only unfreeze the fourth stage for fine-tuning. This approach ensures that the network retains general feature extraction capabilities while adapting to the characteristics of remote sensing data. We implement independent training processes for each type of orthogonal basis function, and the process can be roughly expressed as follows:

[0035] (6)

[0036] in, represents a fixed-parameter visual backbone network (for spatial feature extraction of remote sensing images); Represents the multi-basis function classifier module; It represents the feature fusion of two types of components, namely the complementary combination of the deep spatial features of ConvNeXt and the polynomial basis function features of KANLinear. The parameter freezing strategy is: freeze the ConvNeXt The layer parameters keep the pre-trained features and only the classifier module is trained.

[0037] S22 Configure loss function and optimizer

[0038] Using cross entropy loss as the loss function, the loss function formula is:

[0039] (7)

[0040] in, is the total number of categories, One-hot encoding of the true label; is the predicted probability. The embodiment uses the Adam optimizer. The Adam optimizer parameter update formula is as follows:

[0041] (8)

[0042] in, is the learning rate, is the decay rate of the first and second order moments, are the first and second order moments, is the first and second order moments corrected for the bias, is the smoothing term.

[0043] S23 Determination of training termination conditions

[0044] The best model is triggered based on the performance of the validation set. The basic logic is as follows:

[0045] (9)

[0046] That is, when the verification accuracy Exceeding historical best The training cycle is terminated when . represents the best model, and Represent the current parameter configuration and the historical best parameter configuration respectively. The model training loss and accuracy curve of the embodiment are as follows Figure 4 As shown, it can be seen that the model has converged at this time and has a good classification effect.

[0047] S24 Model Save

[0048] The model that meets the termination condition is considered the optimal model and is saved as the optimal model. The optimal KAN model saving logic for different basis functions is as follows: the basis function KAN model corresponding to Hermite is stored in convnext_kan_Hermite.pth; the basis function KAN model corresponding to Legendre is stored in convnext_kan_Legendre.pth; the basis function KAN model corresponding to Jacobi is stored in convnext_kan_Jacobi.pth

[0049] S3 generates classification results of remote sensing images to be processed

[0050] S31 Data preparation stage

[0051] This stage is characterized by preprocessing the remote sensing images to be processed, mainly including normalization processing and data stream encapsulation.

[0052] The theoretical formula for normalization is as follows:

[0053] (10)

[0054] in, Indicates the channel identifier of the input remote sensing image; and Represents the first and second The original pixel value of the channel; and Represent the preset mean and standard deviation of channel c, respectively. Data stream encapsulation includes encapsulation of the training set and the test set. Both the training set and the test set encapsulation use a batch size of 32.

[0055] S32 model processing stage

[0056] This stage is characterized by the main use of parameter layered loading and basis function switching mechanisms.

[0057] Parameter layered loading follows the formula loading:

[0058] (11)

[0059] in, Indicates model parameter loading, For feature extraction backbone network parameters locked by param.requires_grad=False, The classifier component loaded using torch.load.

[0060] The basis function switching mechanism loads the KAN modules corresponding to Hermite / Laguerre / Jacobi through the configuration file. The specific correspondence is as follows:

[0061] (12)

[0062] in, Represents the KAN module selection function based on the input data features, and different KAN module branches are selected according to the conditions.

[0063] S33 Result output stage

[0064] This stage is characterized by outputting and saving the classification results of remote sensing data. Some classification results of the embodiment are as follows Figure 5 As shown (the left sub-image represents the actual image and category, the middle sub-image represents the probability distribution histogram, and the right sub-image is the predicted category image), it can be seen that the classification model has a good classification effect.

[0065] Beneficial effects

[0066] The Orthogonal-KAN model and adaptive training method of the present invention have the following significant advantages in the field of satellite remote sensing image classification:

[0067] 1. Lightweight Computing Architecture: By freezing the front-end convolutional layer parameters of the pre-trained feature extraction backbone network, the computational overhead during training is significantly reduced while retaining the ability to express common features. Combined with a multi-basis function classification module designed with depthwise separable convolution, this effectively reduces the number of model parameters and adapts to the massive processing requirements of high-resolution satellite remote sensing imagery.

[0068] 2. Orthogonal basis function system: The constructed Hermite / Laguerre / Jacobi orthogonal polynomial basis function classifier supports switching to meet the adaptability requirements in different remote sensing scenarios.

[0069] 3. Stable and efficient training mechanism: Adopting a layered weight initialization strategy, the basis function switching mechanism supports parameter inheritance and sharing across models, improving the efficiency of model iterative optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 : Flowchart of the ConvNeXt-OrthoKAN model satellite remote sensing image classification method provided by the embodiment of the present invention

[0071] Figure 2 : Schematic diagram of the ConvNeXt-OrthoKAN model training process provided by an embodiment of the present invention

[0072] Figure 3 Schematic diagram of the remote sensing image processing process of the ConvNeXt-OrthoKAN model provided by the embodiment of the present invention

[0073] Figure 4 : ConvNeXt-OrthoKAN model training loss and accuracy curve provided by the embodiment of the present invention

[0074] Figure 5 : Example diagram of some classification results of the ConvNeXt-OrthoKAN model provided by the embodiment of the present invention DETAILED DESCRIPTION

[0075] The present invention will then fully and clearly explain the implementation details and technical features of the relevant technical solutions in conjunction with several specific implementation plans shown in the accompanying drawings. It should be clear that the embodiments shown in this specification are only some practical cases of the technical solution of the present invention, and their number and implementation methods do not constitute the entire limitation of the protection scope of this solution. Under the guidance of the technical ideas of the present invention, equivalent implementation plans or improved solutions derived by technical personnel in related fields through conventional technical means all fall within the scope of protection defined by the claims of the patent application.

[0076] S1: Construct the ConvNext-OrthoKAN model. This model consists of a feature extraction module (ConvNext-Tiny) and a multi-basis function classifier module (Orthogonal-KAN). Orthogonal-KAN is a KAN network structure that uses Hermite, Laguerre, and Jacobi orthogonal polynomials as basis functions.

[0077] S2: Model training and generation. The ConvNext-OrthoKAN model is trained using the remote sensing image dataset to obtain the trained ConvNext-OrthoKAN model.

[0078] S3: Generate classification results of the remote sensing image to be processed. Use the trained ConvNext-OrthoKAN model to complete the processing of the remote sensing image to be processed and obtain the classification results of the remote sensing image to be processed.

[0079] S1 Construction of Orthogonal-KAN Model

[0080] The model consists of a feature extraction module and a multi-basis function classifier module, specifically including:

[0081] S11 feature extraction module

[0082] The pre-trained ConvNeXt-Tiny architecture is used as the feature extraction backbone network to extract multi-scale satellite remote sensing image features and freeze the pre-trained parameters to reduce computational overhead. The mathematical representation of feature extraction is:

[0083] (1)

[0084] In formula (1) is the feature extraction function (where subscript Represents Extraction), ConvBlock is depthwise separable convolution stages, the final output dimension is The feature tensor of 、 Indicates the height and width of the image, is the number of output channels of the last convolutional layer of ConvNeXt-Tiny).

[0085] S12 multi-basis function classifier module

[0086] A configurable KAN architecture was designed, using a KAN layer composed of orthogonal polynomial basis functions for nonlinear transformations. Three independent classifier modules were developed, each integrating a different orthogonal polynomial basis function. The polynomials of the different basis functions were subjected to Tanh input normalization. The multi-basis function classifier module uses Hermite / Laguerre / Jacobi orthogonal polynomials as the KAN basis functions, and the classifier output is:

[0087] (2)

[0088] In formula (2) is the highest order, is an orthogonal polynomial basis of set degree (Hermite, Laguerre or Jacobi), The KAN subscript poly indicates that the nonlinear basis function type of the classifier is an orthogonal polynomial, which is used to distinguish KAN classifiers with different basis function types. The step-by-step recursive relationship of different basis functions is as follows:

[0089] Hermite basis functions:

[0090] (3)

[0091] Laguerre polynomials (with parameters ):

[0092] (4)

[0093] Jacobi polynomials (with parameters ):

[0094] (5)

[0095] in, is a parameter that controls the weight distribution of the polynomial. By adjusting The value of can change the local sensitive area of ​​the basis function on the input feature normalization area, making it more adaptable to the statistical characteristics of the remote sensing image features.

[0096] S2 model training and generation

[0097] S21 Initialize ConvNeXt-OrthoKAN

[0098] Using a pre-trained ConvNeXt-Tiny backbone network, we freeze the parameters of the first three stages to maintain the stability of the pre-trained parameters, and only unfreeze the fourth stage for fine-tuning. This approach ensures that the network retains general feature extraction capabilities while adapting to the characteristics of remote sensing data. We implement independent training processes for each type of orthogonal basis function, and the process can be roughly expressed as follows:

[0099] (6)

[0100] in, represents a fixed-parameter visual backbone network (for spatial feature extraction of remote sensing images); Represents the multi-basis function classifier module; It represents the feature fusion of two types of components, namely the complementary combination of the deep spatial features of ConvNeXt and the polynomial basis function features of KANLinear. The parameter freezing strategy is: freeze the ConvNeXt The layer parameters keep the pre-trained features and only the classifier module is trained.

[0101] S22 Configure loss function and optimizer

[0102] Using cross entropy loss as the loss function, the loss function formula is:

[0103] (7)

[0104] in, is the total number of categories, One-hot encoding of the true label; is the predicted probability. The embodiment uses the Adam optimizer. The Adam optimizer parameter update formula is as follows:

[0105] (8)

[0106] in, is the learning rate, is the decay rate of the first and second order moments, are the first and second order moments, is the first and second order moments corrected for the bias, is the smoothing term.

[0107] S23 Determination of training termination conditions

[0108] The best model is triggered based on the performance of the validation set. The basic logic is as follows:

[0109] (9)

[0110] That is, when the verification accuracy Exceeding historical best The training cycle is terminated when . represents the best model, and Represent the current parameter configuration and the historical best parameter configuration respectively. The model training loss and accuracy curve of the embodiment are as follows Figure 4 As shown, it can be seen that the model has converged at this time and has a good classification effect.

[0111] S24 Model Save

[0112] The model that meets the termination condition is considered the optimal model and is saved as the optimal model. The optimal KAN model saving logic for different basis functions is as follows: the basis function KAN model corresponding to Hermite is stored in convnext_kan_Hermite.pth; the basis function KAN model corresponding to Legendre is stored in convnext_kan_Legendre.pth; the basis function KAN model corresponding to Jacobi is stored in convnext_kan_Jacobi.pth

[0113] S3 generates classification results of remote sensing images to be processed

[0114] S31 Data preparation stage

[0115] This stage is characterized by preprocessing the remote sensing images to be processed, mainly including normalization and data stream encapsulation. The theoretical formula for normalization is as follows:

[0116] (10)

[0117] in, Indicates the channel identifier of the input remote sensing image; and Represents the first and second The original pixel value of the channel; and Respectively represent The preset mean and standard deviation of the channel. Data stream encapsulation includes encapsulation of the training set and the test set. The encapsulation of the training set and the test set uses a batch size of 32.

[0118] S32 model processing stage

[0119] This stage is characterized by the main use of parameter layered loading and basis function switching mechanisms.

[0120] Parameter layered loading follows the formula loading:

[0121] (11)

[0122] in, Indicates model parameter loading, For feature extraction backbone network parameters locked by param.requires_grad=False, The classifier component loaded using torch.load.

[0123] The basis function switching mechanism loads the KAN modules corresponding to Hermite / Laguerre / Jacobi through the configuration file. The specific correspondence is as follows:

[0124] (12)

[0125] in Represents the KAN module selection function based on the input data features, and different KAN module branches are selected according to the conditions.

[0126] S33 Result output stage

[0127] This stage is characterized by outputting and saving the classification results of remote sensing data. Some classification results of the embodiment are as follows Figure 5 As shown (the left sub-image represents the actual image and category, the middle sub-image represents the probability distribution histogram, and the right sub-image is the predicted category image), it can be seen that the classification model has a good classification effect.

Claims

1. A satellite remote sensing image classification method based on the ConvNext-OrthoKAN model, which features include The steps are as follows: S1: Constructing a ConvNext-OrthoKAN model; the model consists of a feature extraction module (ConvNext-Tiny) and a multi-basis function classifier module (Orthogonal-KAN). Orthogonal-KAN is a KAN network structure that uses Hermite, Laguerre, and Jacobi orthogonal polynomials as basis functions. S2: Model training and generation: Use remote sensing image datasets to train the ConvNext-OrthoKAN model and obtain the trained ConvNext-OrthoKAN model; S3: Generate classification results of remote sensing images to be processed; use the trained ConvNext-OrthoKAN model to complete the processing of remote sensing images to be processed and obtain classification results of remote sensing images to be processed.

2. The satellite remote sensing image classification method based on the ConvNext-OrthoKAN model according to claim 1, characterized in that: In S1, the pre-trained ConvNeXt-Tiny architecture is used as the feature extraction backbone network to extract multi-scale satellite remote sensing image features, and the pre-trained parameters are frozen to reduce computational overhead; the mathematical representation of feature extraction is: (1) In formula (1) is the feature extraction function (subscript Represents Extraction), ConvBlock is depthwise separable convolution stages, the final output dimension is The feature tensor of 、 Indicates the height and width of the image, is the number of output channels of the last convolutional layer of ConvNeXt-Tiny).

3. The satellite remote sensing image classification method based on the ConvNext-OrthoKAN model according to claim 1, characterized in that: In S1, a configurable KAN architecture is designed. Nonlinear transformation is performed through the KAN layer composed of orthogonal polynomial basis functions. Three independent classifier modules are developed, each integrating different orthogonal polynomial basis functions. The polynomials of different basis functions are processed by Tanh input normalization. The multi-basis function classifier module uses Hermite / Laguerre / Jacobi orthogonal polynomials as the basis functions of KAN. The classifier output is: (2) In formula (2) is the highest order, is an orthogonal polynomial basis of set degree (Hermite, Laguerre or Jacobi), The KAN subscript poly indicates that the nonlinear basis function type of the classifier is an orthogonal polynomial, which is used to distinguish KAN classifiers with different basis function types. The step-by-step recursive relationship of different basis functions is as follows: Hermite basis functions: (3) Laguerre polynomials (with parameters ): (4) Jacobi polynomials (with parameters ): (5) in, is a parameter that controls the distribution of polynomial weights; by adjusting The value of can change the local sensitive area of ​​the basis function on the input feature normalization area, making it more adaptable to the statistical characteristics of satellite remote sensing image features.

4. The satellite remote sensing image classification method based on the ConvNext-OrthoKAN model according to claim 1, characterized in that: In S2, a pre-trained ConvNeXt-Tiny backbone network is used. The parameters of the first three stages are frozen to maintain the stability of the pre-trained parameters, and only the fourth stage is unfrozen for fine-tuning. This solution ensures that the network retains general feature extraction capabilities and can adapt to the characteristics of satellite remote sensing data. Independent training processes are implemented for various orthogonal basis functions. The process can be roughly expressed as follows: (6) in, represents a fixed-parameter visual backbone network (for spatial feature extraction of remote sensing images); Represents the multi-basis function classifier module; It represents the feature fusion of two types of components, namely the complementary combination of the deep spatial features of ConvNeXt and the polynomial basis function features of KANLinear; the parameter freezing strategy is: freeze the ConvNeXt front The layer parameters keep the pre-trained features and only the classifier module is trained.

5. The satellite remote sensing image classification method based on the ConvNext-OrthoKAN model according to claim 1, characterized in that: In S2, cross entropy loss is used as the loss function, and the loss function formula is: (7) in, is the total number of categories, is the one-hot encoding of the true label; is the predicted probability.

6. The satellite remote sensing image classification method based on the ConvNext-OrthoKAN model according to claim 1, characterized in that: In S2, the Adam optimizer is used; the parameter update formula of the Adam optimizer is as follows: (8) in, is the learning rate, is the decay rate of the first and second order moments, are the first and second order moments, is the first and second order moments corrected for the bias, is the smoothing term.

7. The satellite remote sensing image classification method based on the ConvNext-OrthoKAN model according to claim 1, characterized in that: In S2, the best model is triggered based on the performance of the validation set. The basic logic is as follows: (9) That is, when the verification accuracy Exceeding historical best The training cycle is terminated when represents the best model, and Represent the current parameter configuration and the historical best parameter configuration respectively.

8. The satellite remote sensing image classification method based on the ConvNext-OrthoKAN model according to claim 1, characterized in that: In S3, the following three phase steps are performed in sequence: The first stage is the data preparation stage; this stage is characterized by preprocessing the remote sensing images to be processed, mainly including normalization processing and data stream encapsulation; the theoretical formula of normalization processing is as follows: (10) in, Indicates the channel identifier of the input remote sensing image; and Represents the first and second The original pixel value of the channel; and Respectively represent The preset mean and standard deviation of the channel; data stream encapsulation includes the encapsulation of the training set and the test set; the encapsulation of the training set and the test set both use a batch size of 32; The second stage is the model processing stage; this stage is characterized by the use of parameter layered loading and basis function switching mechanism; parameter layered loading follows the formula loading: (11) in, Indicates model parameter loading, For feature extraction backbone network parameters locked by param.requires_grad=False, This is the classifier component loaded using torch.load; the basis function switching mechanism loads the KAN module corresponding to Hermite / Laguerre / Jacobi through the configuration file, specifically as follows: (12) in, Represents the KAN module selection function based on the input data characteristics, and selects different KAN module branches according to the conditions; The third stage is the result output stage; this stage is characterized by outputting and saving the classification results of remote sensing data.

9. The satellite remote sensing image classification method based on the ConvNext-OrthoKAN model according to claim 1, characterized in that: The number of deep separation convolution stages Set to 4, the highest order of orthogonal polynomial Set to 5, set in Laguerre classifier , Jacobi classifier is set , Adam's learning rate Set to 0.05.

Images