Land coverage classification method based on reasoning segmentation

Abstract

The invention discloses a land coverage classification method based on inference segmentation, and relates to the technical field of remote sensing image inference segmentation and deep learning. Comprising the steps of training sample set establishment, multi-scale feature extraction submodule design, cross-modal feature fusion module design, land coverage classification model structure design based on reasoning segmentation, land coverage classification model training based on reasoning segmentation, model performance evaluation and index analysis. According to the method, effective combination of remote sensing image features and semantic information is realized, and the differentiated cognitive ability for different ground object targets is improved. Meanwhile, a remote sensing image land cover intelligent classification technology with practical value is obtained, the professional technical threshold is remarkably reduced, the intelligence and universality of remote sensing image interpretation are realized, and a convenient remote sensing information acquisition method is provided for users in various fields.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image inference segmentation and deep learning, and particularly relates to a land cover classification method based on inference segmentation. BACKGROUND

[0002] Land cover classification is one of the core tasks of remote sensing image interpretation, and has important significance for ecological environment monitoring, urban planning and agricultural production. Through accurate classification of ground objects in remote sensing images, the influence of urban expansion on forest ecosystems can be effectively identified, the flood disaster risk of the surrounding area of wetlands can be evaluated, and important references can be provided for agricultural production planning and food security assessment. Accurate land cover classification results can provide reliable technical support for regional sustainable development decision-making, and are an important basis for realizing intelligent regional management.

[0003] The land cover classification method has undergone a technical evolution process from manual interpretation, machine learning to deep learning. In the first stage, the research personnel mainly relied on artificial feature extraction and used statistical modeling methods to extract information from remote sensing data, including fuzzy clustering, visual interpretation, decision tree and other statistical models for classification. With the advent of the "quantitative remote sensing era", researchers began to develop physical models using global quantitative remote sensing products to study the interaction process and results between remote sensing information and transmission medium and target. Compared with statistical models, physical models can solve more different application problems, but these methods are difficult to be widely applied to different regions and environments due to the large differences in topographic features of different regions. In the second stage, with the development of computer technology, shallow machine learning methods such as support vector machine (SVM) and artificial neural network (BP) were gradually applied to remote sensing classification. Existing research has proposed a classification method based on cost-sensitive active learning to improve classification accuracy by constructing a reliable training set; at the same time, a supervised classification method for multi-spectral satellite images based on detail-enhanced texture feature extraction algorithm is proposed, which eliminates noise by normal contraction technology and calculates the co-occurrence features on the gray level co-occurrence matrix; a new multi-center supervised fuzzy multi-classification method is proposed to solve the spectral diversity problem in multi-spectral remote sensing image classification. However, these shallow learning methods use a small amount of sample data and have a limited number of model parameters, which limits the generality of the model.

[0004] With the development of deep learning technology, methods based on convolutional neural network (CNN) have gradually become mainstream, and remote sensing information extraction has entered the third stage. FCN first introduced CNN into the field of semantic segmentation, providing a new technical paradigm for remote sensing image classification. U-Net effectively fuses high and low level features through a skip connection structure, significantly improving the ability to preserve boundary details. DeepLab series introduces the dilated convolution and ASPP module, enhancing the multi-scale feature extraction capability. Various deep learning-based remote sensing information extraction methods have appeared in existing technologies, including using hyperspectral remote sensing image technology and convolutional neural network for coal mine area feature information extraction, using deep fully convolutional network to predict land cover categories, and high-resolution remote sensing image building extraction based on adaptive pooling model. In recent years, to cope with the dependence of deep learning methods on a large number of labeled samples, new learning methods have emerged, such as metric-based few-shot classification technology, self-supervised pre-training method, and contrastive learning-based self-supervised feature extraction method. In addition, the introduction of the transformer architecture brings new breakthroughs to land cover classification. VisionTransformer divides the image into a patch sequence and uses the self-attention mechanism to establish long-range dependencies. SwinTransformer introduces a hierarchical window attention mechanism, which better adapts to the multi-scale characteristics of remote sensing images. However, existing methods still have some shortcomings: existing deep learning models mainly rely on pixel-level features, lacking a deep understanding of the spatial relationships between features; the surrounding environment and location information of features are not fully utilized, making it difficult to establish a context-based semantic understanding; the diversity of features of the same feature in different spatio-temporal scales and different sensor-acquired remote sensing images can easily lead to cognitive confusion of existing models.

[0005] To solve the above problems, the present application provides a land cover classification method based on reasoning segmentation. This method innovatively combines image features with semantic information, and realizes differentiated cognition of different feature targets by analyzing the relative position, distribution characteristics and surrounding environment of features. This spatial relationship-based semantic understanding method not only improves the accuracy of remote sensing image interpretation, but also provides an intuitive human-computer interaction method, allowing non-professional users to specify target areas of interest through natural language, greatly reducing the threshold for professional applications. At the same time, this method can deeply explore the interaction mechanism between features, providing a more comprehensive analysis perspective for ecological environment monitoring. SUMMARY

[0006] To solve the problems in the above background art, the present application provides a land cover classification method based on reasoning segmentation.

[0007] To achieve the above object, the application is implemented by the following technical solutions: A land cover classification method based on reasoning segmentation, comprising the following steps:

[0008] Step one, establish a training sample set, as follows:

[0009] (1) Obtain high-resolution satellite remote sensing images, cut and process them into 1024*1024 size RGB images, and finally manually label the land cover types, including six types of forest land, grassland, wetland, farmland, town and road;

[0010] (2) Use Python script to process and generate semantic description of the labeled image; first, number each land cover type (forest land, grassland, wetland, farmland, town and road) and create corresponding binary mask; then generate semantic description in three ways: first, based on area analysis, calculate the proportion of target area pixel area to the whole, and divide it into three categories: maximum (10%~100%), medium (5%~10%) and minimum (0.5%~1%); the second is based on position analysis, the image is divided into nine regions (upper, middle, lower, left, right, upper left, upper right, lower left and lower right), and the position description is generated according to the center point position of the target area; the third is based on spatial relationship analysis, by calculating the overlap degree and distance between target class and reference class, to judge whether they are adjacent or surrounded; finally, all processed samples are randomly divided into training set, test set and validation set according to the proportion of 7:2:1, and saved to the corresponding text file;

[0011] (3) Use Python script to generate files required for model training; first, read the text files of training set, test set and validation set to get the image number and description information of each sample; then process each mask image: convert it to binary image, calculate run-length encoding as segmentation information, calculate bounding box coordinates and area features, and write these information into JSON format instance file; finally, create complete reference information for each sample, including reference number, image number, class number, data set division type, semantic description and its segmentation result, etc., and save these information as P file after serialization; the two generated files will be used as input data for model training;

[0012] Step two, design a multi-scale feature extraction submodule, as follows:

[0013] Based on the features of remote sensing image, the input remote sensing image is processed by two key components of cross-scale interaction and non-local filtering; first, use the cross-scale interaction module to process the input features of different scales; the features of each layer are unified to the same size by downsampling operation:

[0014]

[0015] Then these features are spliced in the channel dimension to obtain multi-scale features:

[0016]

[0017] Next, a multi-receptive field feature representation is constructed using a depth separable convolution; the spliced features are subjected to multi-scale transformation:

[0018]

[0019] where k m represents a convolution kernel of different scales; these features are then optimized by a multi-scale attention layer and further enhanced using local relationship compensation:

[0020]

[0021] where DWConv represents a depth separable convolution, and Hardswish is an activation function; at the same time, the non-local filtering module first converts the spatial domain feature e into the frequency domain using a fast Fourier transform:

[0022] E = F[e] ∈ C H×W×C

[0023] where F represents a fast Fourier transform operation, H and W represent the height and width of the feature map respectively, and C represents the number of feature channels; after the conversion, a learnable filter K is used to modulate the feature in the frequency domain to capture global context information:

[0024]

[0025] where ⊙ represents the Hadamard product, and K represents the learnable filter parameter; finally, the modulated frequency domain feature is converted back to the spatial domain through an inverse Fourier transform:

[0026]

[0027] where F -1 represents an inverse fast Fourier transform operation, represents the modulated frequency domain feature, represents the feature converted back to the spatial domain, and R represents a tensor in the real number field; after the operation, the spatial and frequency information is fused through a jump connection, and finally the feature channel dimension is adjusted through a 1×1 convolution; this feature extraction method combining cross-scale interaction and non-local filtering can effectively process multi-scale features in remote sensing images, and can simultaneously obtain local details and global semantic information, significantly improving the model's positioning and recognition ability for targets;

[0028] Step three, cross-modal feature fusion module design, as follows:

[0029] This module first extracts the spatial features of the image through the visual state space layer, and uses a four-way scanning strategy to capture the dependence relationship in different directions; then realizes multi-modal feature fusion in the Twister layer: first, through image-text interaction operation at global and local levels, the similarity of image and text features is calculated and mapped to a shared feature space; among them, the local interaction operation establishes a fine-grained feature correspondence by calculating the similarity matrix of image features and text features; specifically, the local interaction mapping F c The calculation formula is:

[0030] F c = F i W i ·(F t W t ) T ∈R H×W×L

[0031] Where F i represents the input image feature, F t represents the input text feature, W i represents the learnable projection matrix of image features, W t represents the learnable projection matrix of text features, H and W represent the height and width of the feature map respectively, L represents the length of the text sequence, and R represents a tensor in a real number field, which projects image and text features to a shared feature space for similarity calculation through mapping; in order to enhance the processing ability of high-dimensional features, a convolutional layer is used to convert the local interaction feature mapping result to a feature representation of the required dimension; then, the image features, global text features and local interaction features are concatenated in the channel dimension to construct a mixed feature, ensuring that the image features can simultaneously obtain local and global text information guidance; finally, a twist mechanism is designed to enhance the feature interaction within and across modalities through two state space models of continuous channel scanning and spatial scanning, where channel scanning promotes the fusion of different modal features, and spatial scanning learns the spatial dependence within a single modality; the fused features are generated into the final segmentation result through the decoder, and the entire module is trained in an end-to-end manner to realize adaptive learning of the feature fusion strategy.

[0032] Step four, land cover classification model structure design based on reasoning segmentation, as follows:

[0033] The overall structure of the model includes a Twister state space main network, a double-path feature extraction branch, an image-text feature fusion module and an improved decoder structure; the main network is based on the Mamba architecture and uses the Twister module as the core computing unit to capture the long-range dependencies of images through state space modeling and multi-scale feature interaction; the double-path feature extraction branch includes a main branch and an auxiliary branch: the main branch uses Twister state space sequence modeling to extract main features, and the auxiliary branch uses a lightweight ResBlock structure to extract supplementary features; the two branches fuse features through a designed FusionBlock module, which includes an adaptive attention mechanism and position encoding, which can effectively integrate feature information from different sources; the image-text feature fusion module uses an innovative structure design: first, the text feature is projected into the visual space through the text guidance module, then the image-text feature fusion module is used to calculate the correlation between image and text features, and finally the local-global feature fusion is used to enhance the expression ability of the model; the module also includes a multi-scale feature extraction submodule that can integrate and optimize features at different resolutions; the decoder uses a progressive design, uses ResBlock for feature transformation, introduces a non-local filtering module for feature enhancement in the frequency domain, and gradually restores the feature resolution through skip connections; feature optimization modules are provided at each stage of the decoder, including channel attention and spatial attention mechanisms, to ensure that the generated segmentation results retain global information and contain rich local details;

[0034] Step five, land cover classification model training based on inference segmentation, specifically as follows:

[0035] Randomly extract 80% of the data from the training sample set constructed in step one for training, and the remaining 20% of the data for verification; first, input the remote sensing image and the corresponding semantic description in the training set into the network designed in step four; the remote sensing image obtains multi-scale feature representation through the Twister state space main network and the double-path feature extraction branch, while the semantic description is fused with the image features in the cross-modal feature fusion module after text encoding, and finally the progressive decoder generates segmentation prediction results; during the training process, the model performance is evaluated using the validation set after the end of each epoch, and the model weight with the best performance is saved; in order to improve the generalization ability of the model, data augmentation strategies are used during the training process, including random cropping, horizontal flipping, vertical flipping, etc.; at the same time, a learning rate decay strategy is introduced, which reduces the learning rate when the validation set performance does not improve for multiple epochs;

[0036] Step six, model performance evaluation and index analysis, specifically as follows:

[0037] The application verifies the effectiveness of the land cover classification model based on reasoning segmentation, and comprehensively evaluates the performance of the model; detailed experimental analysis is carried out on the data set constructed in step one, and the main evaluation indexes include mean intersection over union (mIoU), overall intersection over union (oIoU) and precision (Precision@X, where X belongs to {50, 60, 70, 90}) under different thresholds; wherein, mIoU is used to evaluate the segmentation accuracy of the model on each class, oIoU reflects the overall segmentation performance, and Precision@X represents the proportion of test samples with IoU score higher than a certain threshold X%;

[0038] At the same time, the calculation efficiency of the model is analyzed, including the inference speed (FPS) and GPU memory occupation; the model performance is tested under different input resolutions, and it is verified that the application has higher efficiency and lower resource consumption compared with the existing architecture; in addition, the fusion ability of the model to the semantic features of the image is qualitatively analyzed through visual attention map and local interaction mapping; we also conduct comparative experiments with the current most advanced method, and comprehensively evaluate the advantages of the application in accuracy and efficiency; the experimental results show that the application significantly reduces the computational complexity while maintaining high segmentation accuracy, and has good practical value.

[0039] The application provides a land cover classification method based on reasoning segmentation, which has the following beneficial effects:

[0040] The land cover classification method based on reasoning segmentation solves the problems of feature cognitive confusion and high professional threshold in current land cover classification by combining reasoning segmentation technology with multi-modal feature fusion to construct a land cover classification model based on reasoning segmentation.

[0041] The method effectively avoids the feature cognitive confusion problem caused by the dependence of traditional deep learning methods on a large number of sample training by introducing semantic description and spatial relationship information, and uses natural language interaction to enable non-professional users to accurately describe and obtain the target area information of interest, greatly reducing the professional technical threshold.

[0042] In practical application, the application not only significantly improves the accuracy of land cover classification, but also realizes the intelligentization and universalization of remote sensing image interpretation, and provides a convenient remote sensing information acquisition method for users in various fields. BRIEF DESCRIPTION OF DRAWINGS

[0043] Fig. 1 The flowchart of the embodiment of the application.

[0044] Fig. 2 The model framework diagram of the embodiment of the application.

[0045] Fig. 3A result display diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments.

[0047] In the description of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more than two; the terms "upper", "lower", "left", "right", "inner", "outer", "front end", "rear end", "head", "tail" and the like indicate the orientation or positional relationship shown in the drawings, which is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", "third" and the like are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0048] In the description of the present application, it should be noted that, unless otherwise specified and limited, the terms "connected", "connected" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0049] Please refer to Figs. 1-3 The present application provides a technical solution: a land cover classification method based on reasoning segmentation, comprising the following steps:

[0050] Step 1, establish a training sample set, specifically as follows:

[0051] (1) Obtain high-resolution satellite remote sensing images, cut and process them into 1024*1024 size RGB images, and finally manually label the land cover types, including six types of forest land, grassland, wetland, farmland, town and road;

[0052] (2) Using Python scripts to process the labeled images and generate semantic descriptions; first, each land cover type (forest, grassland, wetland, farmland, town, road) is numbered and a corresponding binary mask is created; then generate semantic descriptions in three ways: the first is based on area analysis, calculate the target area pixel area proportion of the whole, divide it into three categories: maximum (10%~100%), medium (5%~10%) and minimum (0.5%~1%); The second is based on position analysis, the image is divided into nine regions (upper, middle, lower, left, right, upper left, upper right, lower left, lower right), and the position description is generated according to the center point position of the target area; The third is based on spatial relationship analysis, by calculating the overlap degree and distance between target class and reference class, to judge their relationship is adjacent or surrounded; Finally, all the processed samples are randomly divided into training set, test set and validation set according to the proportion of 7:2:1, and saved to the corresponding text file;

[0053] (3) Use Python scripts to generate files required for model training; first, read the text files of the training set, test set, and validation set to obtain the image number and description information of each sample; then process each mask image: convert it to a binary image, calculate the run-length encoding as the segmentation information, calculate the bounding box coordinates and area features, and write these information into a JSON format instance file; Finally, create complete reference information for each sample, including reference number, image number, class number, data set division type, semantic description and its segmentation results, etc., and save these information as P file after serialization; The two generated files will be used as input data for model training;

[0054] Second, the multi-scale feature extraction submodule is designed as follows:

[0055] Based on the features of remote sensing images, the input remote sensing images are processed through two key components: cross-scale interaction and non-local filtering; First, use the cross-scale interaction module to process the input features of different scales; The features of each layer are Unified to the same size through downsampling operation:

[0056]

[0057] Then concatenate these features in the channel dimension to get multi-scale features:

[0058]

[0059] Next, use depth separable convolution to construct multi-receptive field feature representation; The concatenated features Perform multi-scale transformation:

[0060]

[0061] where k m The features are then optimized by a multi-scale attention layer and further enhanced by local relation compensation:

[0062]

[0063] where DWConv represents a depthwise separable convolution, and Hardswish is an activation function; meanwhile, the non-local filtering module first converts the spatial domain feature e into the frequency domain using fast Fourier transform:

[0064] E = F[e] ∈ C H×W×C

[0065] where F represents the fast Fourier transform operation, H and W represent the height and width of the feature map respectively, and C represents the number of feature channels; after the conversion, a learnable filter K is used to modulate the feature in the frequency domain to capture global context information:

[0066]

[0067] where ⊙ represents the Hadamard product, and K represents the learnable filter parameter; finally, the modulated frequency domain feature is converted back to the spatial domain through inverse Fourier transform:

[0068]

[0069] where F -1 represents the inverse fast Fourier transform operation, represents the modulated frequency domain feature, represents the feature converted back to the spatial domain, and R represents a tensor in the real number field; after the operation, the spatial and frequency information is fused through a jump connection, and finally the feature channel dimension is adjusted through a 1×1 convolution; this feature extraction method combining cross-scale interaction and non-local filtering can effectively process multi-scale features in remote sensing images, and can also obtain local details and global semantic information, significantly improving the model's positioning and recognition ability for targets;

[0070] Step three, cross-modal feature fusion module design, as follows:

[0071] The module firstly extracts the spatial features of the image through the visual state space layer, and adopts a four-way scanning strategy to capture the dependency relationship in different directions; then realizes the multi-modal feature fusion in the Twister layer: firstly, through the image-text interaction operation at the global and local levels, the similarity of image and text features is calculated and mapped to the shared feature space; among them, the local interaction operation establishes the fine-grained feature correspondence by calculating the similarity matrix of image features and text features; specifically, the local interaction mapping F c The calculation formula of F c is as follows:

[0072] F i W i ·(F t W t ) T ∈R H×W×L

[0073] wherein F i represents the input image feature, F t represents the input text feature, W i represents the learnable projection matrix of the image feature, W t represents the learnable projection matrix of the text feature, H and W represent the height and width of the feature map respectively, L represents the length of the text sequence, and R represents a tensor in a real number field, which projects the image and text features into a shared feature space for similarity calculation through mapping; in order to enhance the processing ability of high-dimensional features, a convolutional layer is used to convert the local interaction feature mapping result into a feature representation with the required dimension; then, the image feature, the global text feature and the local interaction feature are concatenated in the channel dimension to construct the mixed feature, ensuring that the image feature can simultaneously obtain the guidance of local and global text information; finally, a twisting mechanism is designed, which enhances the feature interaction within and across modalities through two state space models of continuous channel scanning and spatial scanning, wherein the channel scanning promotes the fusion of different modal features, and the spatial scanning learns the spatial dependency within a single modality; the fused features generate the final segmentation result through the decoder, and the whole module is trained in an end-to-end manner to realize the adaptive learning of the feature fusion strategy;

[0074] Step four, the structure design of the land cover classification model based on reasoning segmentation, specifically as follows:

[0075] The overall structure of the model includes a Twister state space main network, a double-path feature extraction branch, an image-text feature fusion module and an improved decoder structure; the main network is based on the Mamba architecture and uses the Twister module as the core computing unit to capture the long-range dependencies of images through state space modeling and multi-scale feature interaction; the double-path feature extraction branch includes a main branch and an auxiliary branch: the main branch uses Twister state space sequence modeling to extract main features, and the auxiliary branch uses a lightweight ResBlock structure to extract supplementary features; the two branches fuse features through a designed FusionBlock module, which includes an adaptive attention mechanism and position encoding, which can effectively integrate feature information from different sources; the image-text feature fusion module uses an innovative structure design: first, the text feature is projected into the visual space through the text guide module, then the image-text feature fusion module is used to calculate the correlation between image and text features, and finally the local-global feature fusion is used to enhance the expression ability of the model; the module also includes a multi-scale feature extraction submodule that can integrate and optimize features at different resolutions; the decoder uses a progressive design, uses ResBlock for feature transformation, introduces a non-local filtering module for feature enhancement in the frequency domain, and gradually restores the feature resolution through skip connections; feature optimization modules are provided at each stage of the decoder, including channel attention and spatial attention mechanisms, to ensure that the generated segmentation results retain global information and contain rich local details;

[0076] Step five, land cover classification model training based on inference segmentation, specifically as follows:

[0077] Randomly extract 80% of the data from the training sample set constructed in step one for training, and the remaining 20% of the data for verification; first, input the remote sensing image and the corresponding semantic description in the training set into the network designed in step four; the remote sensing image passes through the Twister state space main network and the double-path feature extraction branch to obtain multi-scale feature representation, while the semantic description is fused with the image features in the cross-modal feature fusion module after text encoding, and finally the progressive decoder generates segmentation prediction results; during the training process, the model performance is evaluated using the validation set after the end of each epoch, and the model weight with the best performance is saved; in order to improve the generalization ability of the model, data augmentation strategies are used during the training process, including random cropping, horizontal flipping, vertical flipping, etc.; at the same time, a learning rate decay strategy is introduced, which reduces the learning rate when the performance of the validation set does not improve for multiple epochs in a row;

[0078] Step six, model performance evaluation and index analysis, specifically as follows:

[0079] In order to verify the effectiveness of the land cover classification model based on reasoning segmentation, the performance of the model is comprehensively evaluated; detailed experimental analysis is carried out on the data set constructed in step one, and the main evaluation indexes include the average intersection over union (mIoU), the overall intersection over union (oIoU) and the precision (Precision@X, where X∈{50, 60, 70, 90}) under different thresholds; wherein, mIoU is used to evaluate the segmentation accuracy of the model on each class, oIoU reflects the overall segmentation performance, and Precision@X represents the proportion of test samples with IoU score higher than a certain threshold X%;

[0080] At the same time, we analyze the computational efficiency of the model, including the inference speed (FPS) and GPU memory occupation and other indicators; test the model performance under different input resolutions, verify that the present application has higher efficiency and lower resource consumption compared with the existing architecture; in addition, by visualizing the attention map and local interaction mapping, the fusion ability of the model to the semantic features of the image is qualitatively analyzed; we also conduct comparative experiments with the current most advanced method, comprehensively evaluate the advantages of the present application in accuracy and efficiency; the experimental results show that the present application significantly reduces the computational complexity while maintaining high segmentation accuracy, and has good practical value.

[0081] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A land cover classification method based on inductive segmentation, characterized in that, The method comprises the following steps: Step one, establishing a training sample set; Step two, multi-scale feature extraction submodule design; Step three, cross-modal feature fusion module design; Step four, land cover classification model structure design based on reasoning segmentation; Step five, land cover classification model training based on reasoning segmentation; Step six, model performance evaluation and index analysis.

2. The land cover classification method based on inductive segmentation according to claim 1, characterized in that, In step one, high-resolution satellite remote sensing images are obtained, which are cropped and processed into 1024*1024 size RGB images, and finally manually labeled land cover types, including forest land, grassland, wetland, farmland, town, and road; Python scripts are used to process the labeled images and generate semantic descriptions; first, each land cover type is numbered and a corresponding binary mask is created; then, semantic descriptions are generated in three ways: The first is based on area analysis, which calculates the proportion of target area pixel area to the whole, and divides it into three categories: maximum (10%~100%), medium (5%~10%), and minimum (0.5%~1%); The second is based on location analysis, which divides the image into nine regions and generates location descriptions based on the center point position of the target region; The third is based on spatial relationship analysis, which calculates the overlap and distance between target classes and reference classes to determine their relationship, whether adjacent or surrounded; Finally, all processed samples are randomly divided into training set, test set and validation set according to the ratio of 7:2:1, and saved to the corresponding text files; Python scripts are used to generate files required for model training; first, read the training set, test set, and validation set text files to obtain the image number and description information of each sample; then process each mask image: convert it to a binary image, calculate the run-length encoding as segmentation information, calculate the bounding box coordinates and area features, and write the information into a JSON format instance file; finally, create complete reference information for each sample, including reference number, image number, class number, data set division type, semantic description and its word segmentation result, and save the information as a P file; these two generated files will be used as input data for model training.

3. The land cover classification method based on reasoning segmentation according to claim 1, characterized in that, In step two, based on the features of remote sensing images, the input remote sensing images are processed through cross-scale interaction and non-local filtering two key components; first, use the cross-scale interaction module to process the input features of different scales; features of each layer are concatenated unified to the same size by a down-sampling operation: Then concatenate the features in the channel dimension to get multi-scale features: Then, a multi-receptive field feature representation is constructed using depthwise separable convolution; the features after splicing are multi-scale transformation is performed: where k m denotes different scale convolution kernels; the above features are then optimized by a multi-scale attention layer and further enhanced by local relation compensation: Where DWConv represents a depth separable convolution, and Hardswish is an activation function; at the same time, the non-local filtering module first converts the spatial domain feature e to the frequency domain using fast Fourier transform: E = F[e] e C H×W×C Where F represents the fast Fourier transform operation, H and W represent the height and width of the feature map respectively, and C represents the number of feature channels; after conversion, a learnable filter K is used to modulate the feature in the frequency domain to capture global context information: Where ⊙ represents Hadamard product, and K represents the learnable filter parameters; finally, the modulated frequency domain feature is converted back to the spatial domain through inverse Fourier transform: where F -1 denotes an inverse fast Fourier transform operation, denotes the modulated frequency domain feature, denotes the converted back spatial domain feature, R denotes a real number field tensor; after the operation, the spatial and frequency information is fused through a skip connection, and finally the feature channel dimension is adjusted through a 1x1 convolution.

4. The land cover classification method based on reasoning segmentation according to claim 1, characterized in that, In step three, the module firstly extracts spatial features of the image through the visual state space layer and captures the dependency in different directions by using the four-way scanning strategy; then, in the Twister layer, multi-modal feature fusion is realized: firstly, through image-text interaction operation at global and local levels, the similarity of image and text features is calculated and mapped to a shared feature space; among them, the local interaction operation establishes a fine-grained feature correspondence by calculating the similarity matrix of image features and text features; specifically, the local interaction mapping F c The calculation formula is: F c = F i W i ·(F t W t ) T ∈R H×W×L where F i denotes the input image feature, F t denotes the input text feature, W i denotes the learnable projection matrix of image feature, W t denotes the learnable projection matrix of text feature, H and W represent the height, width of feature map respectively, L represents the length of text sequence, R represents a tensor in real field, which projects image and text features into a shared feature space for similarity calculation; in order to enhance the processing ability of high-dimensional features, a convolutional layer is used to convert the local interaction feature mapping result into a feature representation of the required dimension; then, the image feature, global text feature and local interaction feature are concatenated in the channel dimension to construct a mixed feature, ensuring that the image feature can simultaneously obtain the guidance of local and global text information; finally, a twisting mechanism is designed to enhance the feature interaction within and across modalities through two state space models of continuous channel scanning and spatial scanning, wherein the channel scanning promotes the fusion of different modal features, and the spatial scanning learns the spatial dependency within a single modality; the fused features generate the final segmentation result through the decoder, and the whole module is trained in an end-to-end manner to realize the adaptive learning of the feature fusion strategy.

5. The land cover classification method based on reasoning segmentation according to claim 1, characterized in that, In step four, the overall structure of the model includes a Twister state space backbone network, a dual-path feature extraction branch, an image-text feature fusion module, and an improved decoder structure. The backbone network is based on the Mamba architecture and uses the Twister module as the core computing unit. It captures the long-range dependencies of images through state space modeling and multi-scale feature interaction. The dual-path feature extraction branch includes a main branch and an auxiliary branch. The main branch uses Twister state space sequence modeling to extract main features, while the auxiliary branch uses a lightweight ResBlock structure to extract supplementary features. The two branches are fused through a designed FusionBlock module, which includes an adaptive attention mechanism and position encoding, effectively integrating feature information from different sources. The image-text feature fusion module uses an innovative structure design. First, it projects text features into the visual space through a text-guided module. Then, it calculates the correlation between image and text features using an image-text feature fusion module. Finally, it enhances the model's expression ability through local-global feature fusion. The module also includes a multi-scale feature extraction submodule that integrates and optimizes features at different resolutions. The decoder uses a progressive design, with ResBlock for feature transformation, non-local filtering module for feature enhancement in the frequency domain, and jump connection for gradually restoring feature resolution. Each stage of the decoder is equipped with a feature optimization module, including channel attention and spatial attention mechanisms, to ensure that the generated segmentation results retain global information and contain rich local details.

6. The land cover classification method based on reasoning segmentation according to claim 1, characterized in that, In step five, 80% of the data from the training sample set constructed in step one is randomly extracted for training, and the remaining 20% is used for validation. First, the remote sensing images and corresponding semantic descriptions in the training set are input into the network designed in step four. The remote sensing images are processed through the Twister state space backbone network and the dual-path feature extraction branch to obtain multi-scale feature representations. Meanwhile, the semantic descriptions are encoded and fused with the image features in the cross-modal feature fusion module. Finally, the progressive decoder generates segmentation prediction results. During training, the model performance is evaluated using the validation set after each epoch, and the model weights with the best performance are saved. To improve the model's generalization ability, data augmentation strategies are used during training, including random cropping, horizontal flipping, and vertical flipping. At the same time, a learning rate decay strategy is introduced to reduce the learning rate when the validation set performance does not improve for multiple epochs.

7. The land cover classification method based on reasoning segmentation according to claim 1, characterized in that, In step six, the effectiveness of the land cover classification model based on reasoning segmentation is verified, and the performance of the model is comprehensively evaluated. Experimental analysis is performed on the dataset constructed in step one, and the evaluation indicators include mean intersection over union (mIoU), overall intersection over union (oIoU), and precision at different thresholds (Precision@X, where X is an element of {50, 60, 70, 90}). mIoU is used to evaluate the segmentation accuracy of the model for each class, oIoU reflects the overall segmentation performance, and Precision@X represents the proportion of test samples with an IoU score higher than a certain threshold X%. At the same time, the computational efficiency of the model is analyzed, including inference speed (FPS) and GPU memory occupation indicators. The performance of the model is tested under different input resolutions to verify that the present application has higher efficiency and lower resource consumption compared to existing architectures. In addition, the fusion ability of the model for image semantic features is qualitatively analyzed through visual attention maps and local interactive mapping. Comparative experiments are also conducted with the current most advanced methods to comprehensively evaluate the advantages of the above methods in terms of accuracy and efficiency. The experimental results show that the above method significantly reduces the computational complexity while maintaining high segmentation accuracy, and has good practical value.

Images