Remote sensing image land classification method based on SAM multi-order fine tuning

Abstract

The invention belongs to the technical field of remote sensing semantic segmentation information extraction, and particularly relates to a remote sensing image land classification method based on SAM multi-order fine tuning. According to the method, remote sensing multi-classification precision and generalization ability can be improved. Firstly, a sample data set is constructed through visual interpretation based on domestic high-resolution series images; on the basis, a network is designed by utilizing the strong feature extraction capability of an SAM encoder, a domain offset adaptation module and a deep low-rank adaptation module are designed, a mask generation part of an original SAM is improved, and three-stage adaptation fine adjustment is performed on the SAM, so that the feature extraction capability of the network on a remote sensing image is enhanced, and the model performance is improved. In application, the remote sensing land utilization coverage multi-classification technology with practical value is obtained so as to practically promote application and development of a large model in remote sensing image semantic segmentation multi-classification.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing semantic segmentation information extraction, and in particular relates to a remote sensing image land classification method based on SAM multi-order fine-tuning. Background Art

[0002] The extraction of multi-classification information of land use / cover can provide more comprehensive data support for various types of geological research tasks, including environmental monitoring, change detection, land cover mapping and other practical applications. With the rapid development of remote sensing technology, the land cover classification method based on remote sensing images has become the main means of conducting surface resource and environmental surveys and changes. With the advancement of earth satellite observation technology and the popularization of high-resolution remote sensing images, rich surface land use information is provided. High-resolution satellite data has become the main data source in land use multi-classification tasks, and land use multi-classification research based on high-resolution remote sensing images is of great significance.

[0003] The emergence of deep learning technology has brought many excellent and problem-specific methods in the field of land use / cover multi-classification and has become the mainstream method for high-resolution remote sensing image interpretation.

[0004] The advantages of using rotating target detection to identify objects in remote sensing images are obvious. This is because objects in remote sensing images have characteristics such as arbitrary positions, arbitrary directions, small and dense, and the rotating frame in any direction can fit and accurately locate the object without causing a large number of overlaps in different object detection frames.

[0005] However, due to the influence of the variability of remote sensing imaging conditions and the incomplete consistency of the spectral characteristics of the same target in different scenes, the current deep learning-based methods to achieve high-precision and strong generalization remote sensing land use / cover classification usually rely on a large amount of manual annotation and have high coverage diversity. Secondly, due to the lack of analysis of the differences between natural images and remote sensing images, many deep learning-based classification methods have problems with weak feature extraction or overfitting of remote sensing data. Therefore, limited by the available sample data and model feature extraction capabilities, the current deep learning-based land use / cover classification methods have weak generalization capabilities in a wide range of effective applications across a variety of high-resolution remote sensing scenes, and it is often difficult to achieve ideal results when the study area is changed.

[0006] The semantic segmentation model SAM (SegmentAnythingModel) has a powerful feature extraction capability and provides a new idea for improving the generalization ability and accuracy of the land use multi-classification model. Currently, researchers are trying to apply SAM to the remote sensing field using various fine-tuning strategies used in the NLP field, showing excellent performance. These methods often fail to fully consider the significant differences between remote sensing images and the natural images used by SAM, and their feature offset has a certain impact on the performance of the model.

[0007] Therefore, in order to achieve accurate recognition of objects in remote sensing images, the present invention proposes a remote sensing image land classification method based on SAM multi-order fine-tuning. Summary of the invention

[0008] The purpose of the present invention is to provide a remote sensing image land classification method based on SAM multi-order fine-tuning, to solve the problem of weak generalization ability of current mainstream deep learning methods and the problem of feature offset caused by the significant difference between remote sensing images and natural images used by SAM that mainstream methods based on SAM applied to the remote sensing field fail to fully consider, thereby improving the classification accuracy and generalization ability of remote sensing images.

[0009] In order to achieve the above technical objectives and the above technical effects, the present invention is implemented through the following technical solutions:

[0010] The present invention provides a remote sensing image land classification method based on SAM multi-order fine-tuning, comprising the following steps:

[0011] Step 1: Construct a remote sensing land use / cover multi-classification sample dataset;

[0012] Step 2: Construction of remote sensing image land use / cover model based on SAM multi-order fine-tuning;

[0013] Step 3: Model implementation and training;

[0014] Step 4: Model testing and evaluation;

[0015] Furthermore, in step 1, a sample library is prepared based on remote sensing images; first, a sample data set is constructed through visual interpretation based on various remote sensing images according to the requirements of the corresponding land use / cover classification system. At the same time, the present invention is also compatible with existing public remote sensing multi-classification data sets. For the sample data sets used, the present invention is compatible with any band image input, not limited to RGB three-channel data.

[0016] Furthermore, in step 2, the model mainly includes a domain distribution shift adaptation module (DDSA), a depthwise low-rank adaptation module (DLRA), and a mask decoding structure improved by semantic focused adaptation.

[0017] Furthermore, before the image is input into SAM, it is first fine-tuned using the domain shift adaptation module to separate the spectral and spatial domains of the input high-resolution image, and spatial features are extracted independently in each feature dimension through spatial convolution; then, information in each dimension is collected through point-by-point spectral convolution to capture the correlation between different channels; the DDSA module plays a key role in flexibly adapting to the feature distinction between remote sensing images and natural images and preliminary feature extraction. This ensures that the network balances feature extraction, parameters, and computational costs.

[0018] Furthermore, the features fine-tuned by the domain shift adaptation module are input into the image encoder of SAM. Using SAM's ViT-b as the backbone network, the deep low-rank adaptation module extracts feature information from the specific domain based on LoRA and inserts it into the ViT block of the SAM encoder, achieving efficient model fine-tuning.

[0019] Furthermore, LoRA decomposes important features by calculating the rank of the matrix, so that high-dimensional features can be represented by a low-rank matrix; since high-resolution remote sensing images often have more complex features, a large number of parameters make the model too complex, which can easily lead to overfitting and excessive computational burden. At the same time, limited by the fundamental limitations of VIT, LoRA lacks a local receptive field for details, which can easily lead to information loss in the process of projecting it to a low rank and projecting it back. To this end, the inventors designed DLRA, adding a detail convolution branch based on the LoRA layer, and providing additional adjustment parameters during the fine-tuning process. The specific implementation formula is as follows:

[0020] y=W0x+ΔWx=W0x+BAx+C(x) (1)

[0021]

[0022] Among them, the original weight W∈R(din×dout), din and dout are the dimensions of input and output, A∈R(din×r), B∈R(din×r), r<<min{din,dout}; for the input feature x∈R(b×s×din), the output feature y∈R(b×s×dout); f δRepresents a composite function containing Relu and batchnorm, and represents 3×3 and 9×9 convolutions respectively.

[0023] Furthermore, in the model architecture, the constructed DLRA is applied to each transformer module in the image encoder, freezing the image encoder pre-trained weights W to ensure that the prediction results of the image encoder derived from SAM remain consistent before and after fine-tuning; the weight matrix A is randomly initialized using the Kaming method, and the weight matrix B is initialized to zero;

[0024] First, the high-dimensional features extracted by the image encoder transformer layer are initially projected into three downward components: Q for "query", K for "keyword", and V for "value". Then, the weights are updated using a deep low-rank adaptation module. The fine-tuning strategy for each weight parameter is as follows:

[0025]

[0026] Q=XW' q =XW q +dropout(XA q B q +W C.q ) (4)

[0027] V=XW' v =XW v +dropout(XA v B v +W C.v ) (5)

[0028] K=XW k +dropout(W C.k ) (6)

[0029] Among them, W q , W v and W k is the projection of the pre-trained W, whose parameter weights are fixed and cannot be changed; A q , A v , B q and B v is the weight after low-rank decomposition, W C.q , W C.v and W C.k It is the weight updated by the detail convolution branch in DLRA.

[0030] Furthermore, the mask decoder is provided with an SFA branch; specifically, an additional semantic head is added to the mask decoder and the segmentation head is modified to output a multi-class segmentation mask; the segmentation head branch is integrated into the mask decoder to strengthen the representation of high-level features and improve the boundary accuracy in the segmentation mask; in this branch, the neck features of the image embedding output are upsampled four times and directly applied to the multi-class segmentation output.

[0031] Furthermore, in step 3, all experiments were performed using the lightest version of vit-bSAM implemented on PyTorch 1.12.1; a Tesla V100-PCIE 16G memory GPU was used for model training; the size of the dataset was cropped to 1024 × 1024, the batch size and number of model training rounds were set to 1 and 60 respectively, and the model weights with the best accuracy were selected for testing; all compared network methods used equally divided training and validation test datasets; the model used the AdamW optimizer and set the learning rate to 1e-3 to train all models.

[0032] Furthermore, in step 4, the constructed dataset uses high-resolution series images with a resolution of 1 to 2 m and Ziyuan-3 satellite images to produce vectorized labels through manual visual interpretation by experts, forming a comprehensive classification sample set for model training and test evaluation; the model is tested using the divided test set, which specifically includes 7 land use types: forest, meadow, wetland, farmland, residential land, road and other unused land; the datasets are cropped to a size of 1024×1024, and the training images and test images are divided into a ratio of 6:1; finally, the overall accuracy (OA), mean F1 score (mF1), intersection over union (IoU) and mean IoU (mIoU) are used as evaluation indicators to evaluate the results on all foreground categories.

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

[0034] The present invention provides a remote sensing image land classification method based on SAM multi-order fine-tuning. Aiming at the problems that there are many types of objects in remote sensing images, the distribution of large, medium and small targets is uneven, the target scale changes dramatically, there are sometimes small and dense targets, the scale of most targets is small, the detailed information provided is limited, and the target background is complex, etc., with the improvement of remote sensing target detection accuracy as the core goal, a targeted deep learning algorithm is designed to achieve high-precision and automatic target detection under complex conditions of remote sensing images, and improve the remote sensing target detection accuracy. In terms of application, a remote sensing target detection technology with practical value is obtained, in order to effectively promote the application and development of object monitoring in remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0036] Figure 1 is a flow chart of the method of the present invention;

[0037] Figure 2 It is a module schematic diagram of DDSA in the present invention;

[0038] Figure 3 It is a schematic diagram of the module of DLRA in the present invention;

[0039] Figure 4 It is a module schematic diagram of the SFA in the invention;

[0040] Figure 5 It is a model structure diagram of the method of the present invention;

[0041] Among them, Spectral Depthwise: spectral depth convolution, used to extract spectral features; SpatialPointwise: spatial pointwise convolution, used to extract spatial features; n×nConv: n×n convolution; Relu: Relu layer; Bn: Batchnorm layer; Map×3: features with 3 channels; din: input feature dimension; dout: output feature dimension; Trainable: model trainable part; Frozen: model frozen part; Elementwise addition: element-wise addition; Linear: linear layer; Dropout: dropout operation; MLP: fully connected layer; Upsample: upsampling operation; MulticlassMasks: multi-class mask;

[0042] Figure 6 This is a diagram of the classification results of the method of the present invention. DETAILED DESCRIPTION

[0043] 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.

[0044] See also Figure 1-Figure 6, this embodiment provides a remote sensing image land classification method based on SAM multi-order fine-tuning. This method aims at the problems that the land use category features in remote sensing images are complex, the generalization ability and accuracy are restricted, and the existing methods often lack the ability to be widely applied and effectively implemented after changing the research area, research scene and different sensor images. At the same time, although the existing methods based on SAM application and remote sensing fields show relatively excellent generalization ability, they often fail to fully consider the feature offset problem caused by the significant difference between remote sensing images and natural images in the NLP field. The method of the present invention can improve the accuracy and generalization ability of remote sensing multi-classification. This method first constructs a sample data set through visual interpretation based on a domestic high-resolution series of images. On this basis, the network is designed using the powerful feature extraction ability of the SAM encoder, the domain offset adaptation module and the deep low-rank adaptation module are designed, and the mask generation part of the original SAM is improved. The SAM is fine-tuned in three stages to enhance the network's feature extraction ability for remote sensing images and improve model performance. In terms of application, the present invention obtains a remote sensing land use coverage multi-classification technology with practical value, in order to effectively promote the application and development of large models in remote sensing image semantic segmentation multi-classification.

[0045] The specific steps include:

[0046] Step 1: Construct a remote sensing land use / cover multi-classification sample dataset

[0047] This embodiment produces a sample library based on remote sensing images. First, based on the corresponding land use / cover classification system requirements, a sample data set is constructed through visual interpretation based on various remote sensing images. At the same time, this embodiment is also compatible with existing public remote sensing multi-classification data sets. For the sample data sets used, this embodiment can be compatible with any band image input, not limited to RGB three-channel data.

[0048] Step 2: Construction of remote sensing image land use / cover model based on SAM multi-level fine-tuning

[0049] In view of the feature offset problem caused by the significant difference between the high-resolution remote sensing image used in this embodiment and the natural image used by SAM, the feature extraction capability of the SAM encoder is fully utilized. This embodiment combines the deep separable convolution and low-rank adaptation fine-tuning strategy for improvement, and proposes a remote sensing image land use / cover method based on SAM multi-order fine-tuning. This method designs a domain distribution shift adaptation module (DDSA) and a depthwise low-rank adaptation module (DLRA), and adds a semantic focused adaptation branch (SFA) on the basis of retaining most of the structure of the SAM decoder to enhance the model's ability to extract complex features of remote sensing images. Through the designed DDSA, DLRA, and SFA three-order fine-tuning, SAM is applied to the multi-classification field of land use / cover.

[0050] Domain offset adaptation module: This module is mainly used to address the feature offset problem between remote sensing images and natural images used by SAM. Figure 2 As shown in the figure, before the image is input into SAM, it is first fine-tuned by the domain shift adaptation module, which separates the spectral and spatial domains of the input high-resolution image. First, spatial features are extracted independently in each feature dimension through spatial convolution, and then the information of each dimension is collected through point-by-point spectral convolution to capture the correlation between different channels. This module plays a key role in flexibly adapting to the feature distinction between remote sensing images and natural images and preliminary feature extraction. This ensures that the network balances feature extraction, parameters and computational costs.

[0051] Deep low-rank adaptation module: This module is built based on the LoRA fine-tuning strategy and deep separable convolution to achieve efficient fine-tuning of the SAM encoder parameters. SAM uses a MAE pre-trained VisionTransformer (ViT) as the image encoder, and then uses large-scale supervised training. This embodiment uses SAM's ViT-b as the backbone network, and inputs the features fine-tuned by DDSA into SAM's image encoder. LoRA decomposes important features by calculating the rank of the matrix, so that high-dimensional features can be represented by a low-rank matrix. Since high-resolution remote sensing images often have more complex features, a large number of parameters make the model too complex, which can easily lead to overfitting and excessive computational burden. At the same time, limited by the fundamental limitations of VIT, LoRA lacks a local receptive field for details, which can easily lead to information loss in the process of projecting it to a low rank and projecting it back. To this end, the inventors designed DLRA, adding a detail convolution branch based on the LoRA layer to provide additional adjustment parameters during the fine-tuning process. The specific implementation formula is as follows:

[0052] y=W0x+ΔWx=W0x+BAx+C(x) (1)

[0053]

[0054] Among them, the original weight W∈R(din×dout), din and dout are the dimensions of input and output, A∈R(din×r), B∈R(din×r), r<<min{din,dout}; for the input feature x∈R(b×s×din), the output feature y∈R(b×s×dout); f δ Represents a composite function containing Relu and batchnorm, and represents 3×3 and 9×9 convolutions respectively.

[0055] In the model architecture, the constructed DLRA is applied to each transformer module in the image encoder, freezing the image encoder pre-trained weights W, ensuring that the prediction results of the image encoder derived from SAM remain consistent before and after fine-tuning. The inventors use the Kaming method to randomly initialize the weight matrix A and the weight matrix B to zero. First, the high-dimensional features extracted by the image encoder transformer layer are initially projected into three downward components: Q (query), K (keyword), and V (value). Then, the inventors use the DLRA strategy to update the weights. The fine-tuning strategies for each weight parameter are as follows:

[0056]

[0057] Q=XW' q =XW q +dropout(XA q Bq +W C.q ) (4)

[0058] V=XW' v =XW v +dropout(XA v B v +W C.v ) (5)

[0059] K=XW k +dropout(W C.k ) (6)

[0060] Among them, W q , W v and W k is the projection of the pre-trained W, whose parameter weights are fixed and cannot be changed; A q , A v , B q and B v is the weight after low-rank decomposition, W C.q , W C.v and W C.k It is the weight updated by the detail convolution branch in DLRA.

[0061] Semantic Focus Adaptation Module: In the mask stage, a semantic focus adaptation branch is designed for the mask decoder part to accurately map the features extracted by the encoder to the corresponding land use / cover foreground categories. Its structure is as follows Figure 4 Specifically, the inventors added an additional semantic head to the SAM original mask decoder and modified the segmentation head to output a multi-class segmentation mask. The segmentation head branch is integrated into the mask decoder to strengthen the representation of high-level features and improve the boundary accuracy in the segmentation mask. In this branch, the neck features of the image embedding output are upsampled four times and directly applied to the multi-class segmentation output.

[0062] Step 3: Model implementation and training

[0063] In terms of the experimental software environment configuration, the experiments in this embodiment are all operated on the Windows operating system. The network model in this embodiment and the model used in the comparative experiment are based on PyTorch 1.12.1 as the deep learning framework, and the deep learning environment for model training is configured by Anaconda. The inventor used TeslaV100-PCIE 16G memory GPU for model training. The size of each data set was cropped to 1024×1024, the model batch size and the number of model training rounds were set to 1 and 60 respectively, and the model weights with the best accuracy were selected for testing. All comparative network methods use equally divided training and validation test data sets. The model uses the AdamW optimizer and sets the learning rate to 1e-3 to train all models.

[0064] Step 4: Model testing and evaluation

[0065] The dataset constructed in this embodiment uses high-resolution series images with a resolution of 1 to 2 m and Ziyuan-3 satellite images to produce vectorized labels through manual visual interpretation by experts, forming a comprehensive classification sample set for model training and test evaluation. The model was tested using a partitioned test set, which specifically includes seven land use types: forest, meadow, wetland, farmland, residential land, road and other (unused land). The dataset is divided into 1024×1024 size, and the training images and test images are divided in a ratio of 6:1. Finally, the overall accuracy (OA), mean F1 score (mF1), intersection over union (IoU) and mean IoU (mIoU) are used as evaluation indicators to evaluate the results on all foreground categories.

[0066] Among them, step one is to collect and process remote sensing target detection samples; step two provides a specific construction method of a remote sensing image rotation target detection model based on SAM multi-feature fusion, which is the focus of the patent of this embodiment; step three is mainly the implementation of the model of this embodiment and the model algorithm training content; step four tests and evaluates the accuracy of the trained model.

[0067] A specific application of this embodiment is to use the constructed data set for model training and evaluation. During the training process, the sam-b weights provided by SAM are used, the Image-encoder part is used as the backbone, and the rest of the backbone network except the adapter is frozen. As shown in Table 1 and Figure 6 As shown in the figure, the target detection method proposed in this embodiment achieves advanced performance, and each category achieves the best classification results, which is better than other methods. Figure 6 In the test images with large differences, the method of this embodiment benefits from the powerful feature extraction ability of SAM and shows excellent generalization.

[0068] Table 1 Comparison of results of different model methods

[0069]

[0070]

[0071] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A remote sensing image land classification method based on SAM multi-order fine-tuning, characterized in that: The following steps are involved: Step 1: Construct a remote sensing land use / cover multi-classification sample dataset; Step 2: Construction of remote sensing image land use / cover model based on SAM multi-order fine-tuning; Step 3: Model implementation and training; Step 4: Model testing and evaluation.

2. The remote sensing image land classification method based on SAM multi-order fine-tuning according to claim 1 is characterized in that: In step one, a sample library is created based on remote sensing images. First, according to the requirements of the corresponding land use / cover classification system, a sample data set is constructed through visual interpretation based on various remote sensing images.

3. The remote sensing image land classification method based on SAM multi-order fine-tuning according to claim 1 is characterized in that: In step 2, the model mainly includes a domain shift adaptation module, a deep low-rank adaptation module, and a mask decoding structure improved by semantic focus adaptation.

4. The remote sensing image land classification method based on SAM multi-order fine-tuning according to claim 3 is characterized in that: Before inputting the image into SAM, the domain shift adaptation module is first used to fine-tune it, which separates the spectral and spatial domains of the input high-resolution image, and extracts spatial features independently in each feature dimension through spatial convolution; Then, the information of each dimension is collected through point-wise spectral convolution to capture the correlation between different channels.

5. The remote sensing image land classification method based on SAM multi-order fine-tuning according to claim 4 is characterized in that: The features fine-tuned by the domain shift adaptation module are input into the image encoder of SAM. Using SAM's ViT-b as the backbone network, the deep low-rank adaptation module extracts feature information from the specific domain based on LoRA and inserts it into the ViT block of the SAM encoder, achieving efficient model fine-tuning.

6. The remote sensing image land classification method based on SAM multi-order fine-tuning according to claim 5 is characterized in that: LoRA decomposes important features by calculating the rank of the matrix, so that high-dimensional features can be represented by low-rank matrices. Adding detail convolution branches on the basis of the LoRA layer provides additional adjustment parameters during fine-tuning. The specific implementation formula is as follows: y=W0x+ΔWx=W0x+BAx+C(x) (1) Among them, the original weight W∈R(din×dout), din and dout are the dimensions of input and output, A∈R(din×r), B∈R(din×r), r<<min{din,dout}; for the input feature x∈R(b×s×din), the output feature y∈R(b×s×dout); f δ Represents a composite function containing Relu and batchnorm, and represents 3×3 and 9×9 convolutions respectively.

7. The remote sensing image land classification method based on SAM multi-order fine-tuning according to claim 6 is characterized in that: In the model architecture, the constructed DLRA is applied to each transformer module in the image encoder, freezing the image encoder pre-trained weights W to ensure that the prediction results of the image encoder derived from SAM remain consistent before and after fine-tuning; the weight matrix A is randomly initialized using the Kaming method, and the weight matrix B is initialized to zero; First, the high-dimensional features extracted by the image encoder transformer layer are initially projected into three downward components: Q for "query", K for "keyword", and V for "value". Then, the weights are updated using the deep low-rank adaptation module. The fine-tuning strategy for each weight parameter is as follows: Q=XW’ q =XW q +dropout(XA q B q +W C.q ) (4) V=XW' v =XW v +dropout(XA v B v +W C.v ) (5) K=XW k +dropout(W C.k ) (6) Among them, W q , W v and W k is the projection of the pre-trained W, whose parameter weights are fixed and cannot be changed; A q , A v , B q and B v is the weight after low-rank decomposition, W C.q , W C.v and W C.k It is the weight updated by the detail convolution branch in DLRA.

8. The remote sensing image land classification method based on SAM multi-order fine-tuning according to claim 7 is characterized in that: The mask decoder is set with an SFA branch; specifically, an additional semantic head is added to the mask decoder and the segmentation head is modified to output a multi-class segmentation mask; the segmentation head branch is integrated into the mask decoder to strengthen the representation of high-level features and improve the boundary accuracy in the segmentation mask; in this branch, the neck features of the image embedding output are upsampled four times and directly applied to the multi-class segmentation output.

9. The remote sensing image land classification method based on SAM multi-order fine-tuning according to claim 8 is characterized in that: In step 3, all experiments were conducted using the lightest version of vit-b SAM implemented on PyTorch 1.12.1; a Tesla V100-PCIE 16G memory GPU was used for model training; the size of the dataset was cropped to 1024×1024, the batch size and number of model training rounds were set to 1 and 60 respectively, and the model weight with the best accuracy was selected for testing; All compared network methods use equally divided training and validation test datasets; the models are trained using the AdamW optimizer and the learning rate is set to 1e-3 for all models.

10. The remote sensing image land classification method based on SAM multi-order fine-tuning according to claim 9 is characterized in that: In step 4, the constructed dataset uses high-resolution series images with a resolution of 1 to 2 m and Ziyuan-3 satellite images to produce vectorized labels through manual visual interpretation by experts, forming a comprehensive classification sample set for model training and test evaluation; the model is tested using the divided test set, which includes seven land use types: forest, meadow, wetland, farmland, residential land, road and other unused land; the datasets are cropped to a size of 1024×1024, and the training images and test images are divided into a ratio of 6:1; finally, the overall accuracy, average F1 score, intersection over union and average IoU are used as evaluation indicators to evaluate the results on all foreground categories.

Images