Land utilization classification method based on low-frequency characteristics and related equipment

Abstract

The invention relates to the technical field of land utilization classification, and provides a low-frequency feature-based land utilization classification method and related equipment, and the method comprises the steps: obtaining a plurality of target remote sensing images; obtaining low-frequency features of each target remote sensing image by using the land utilization segmentation model, performing continuous decomposition and decoding on each low-frequency feature to obtain final decoding features of each target remote sensing image, and performing semantic segmentation on each final decoding feature to obtain a land utilization classification probability graph of each target remote sensing image; training a land segmentation model based on all land utilization classification probability graphs to obtain a trained land segmentation model; and segmenting a to-be-classified remote sensing image by using the trained land segmentation model to obtain a land utilization classification result of the to-be-classified remote sensing image. According to the method, the accuracy of land utilization classification can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of land use classification, and in particular to a land use classification method based on low-frequency features and related equipment. Background Art

[0002] Land is a fundamental resource for human survival and development. It is not only a crucial component of natural resources but also a crucial vehicle for socioeconomic activities. Land use is the process by which humans meet their needs through specific activities based on the characteristics of the land. Scientifically classifying land use types and clarifying their specific meanings not only reflects changes in natural surface conditions but also reveals the profound impact of human activities on land cover. This is of great significance for ecological and environmental protection, economic development, and social progress. As a core topic in Earth observation, land use research plays a crucial guiding role in disaster prevention and mitigation, urban and rural planning, environmental governance, and food production. In particular, accurate land use classification provides a scientific basis for optimizing resource allocation and promoting sustainable development, thereby better promoting harmonious coexistence between humans and nature.

[0003] With the continuous advancement of remote sensing technology, the spatiotemporal resolution of remote sensing imagery has been further improved. This rich image information has driven the continuous development of refined land use classification methods, which have undergone four major phases: manual visual interpretation, pixel-based classification methods, object-based classification methods, and deep learning-based classification methods. The earliest land use classification relied on manual visual interpretation, a method based on the interpreter's professional judgment and exhibiting high classification accuracy. However, this method was inefficient and could not meet the practical needs of large-scale and rapid classification. With the widespread adoption of computer technology, automated classification methods have gradually emerged and replaced traditional manual interpretation. Pixel-based classification methods, such as maximum likelihood classification, support vector machines, and K-means clustering, utilize the features of individual pixels for classification. While they can achieve a certain degree of classification accuracy, they struggle to effectively exploit spatial relationships between pixels and have limited ability to distinguish complex landforms. Object-oriented classification methods, such as multi-scale segmentation and rule-based classification, segment the image into objects with similar features and classify them based on properties such as texture and shape. Compared to pixel-based methods, object-oriented methods are more able to leverage spatial relationships and contextual information, resulting in improved classification accuracy. However, its segmentation results are easily affected by subjective factors and complex surface features, especially in areas with fuzzy boundaries, the accuracy of target extraction is still limited.

[0004] With the rapid development of artificial intelligence (AI) technology, deep learning-based classification methods have gradually become mainstream tools in land use classification. These methods include convolutional neural networks, Transformer-based segmentation networks, and segmentation networks based on the state-space model Mamba architecture. They can automatically learn abstract, nonlinear, high-level semantic features and demonstrate significant advantages in processing large-scale data and complex land use types. However, most current deep learning-based methods rely primarily on extracting spatial domain features, paying less attention to the potential of frequency domain information. This makes them limited in their ability to handle complex background noise and makes them inadequate for remote sensing imagery with complex backgrounds. A few studies have attempted to improve the representation of complex land use by fusing features from both the spatial and frequency domains. However, these methods often overlook the differentiated nature of high-frequency and low-frequency features in the frequency domain, resulting in the generation of numerous redundant features and difficulty in achieving deep fusion of frequency and spatial domain features. The resulting features contain little information, making it difficult to fully reflect land use types in an area, resulting in low land use classification accuracy. Summary of the Invention

[0005] The present application provides a land use classification method and related equipment based on low-frequency features, which can solve the problem of low accuracy of land use classification.

[0006] In a first aspect, an embodiment of the present application provides a land use classification method based on low-frequency features, the land use classification method comprising:

[0007] Acquire multiple target remote sensing images;

[0008] The land use segmentation model is used to obtain the low-frequency features of each target remote sensing image, and each low-frequency feature is continuously decomposed and decoded to obtain the final decoded features of each target remote sensing image. Each final decoded feature is semantically segmented to obtain the land use classification probability map of each target remote sensing image. The land use classification probability map is used to describe the probability that the area corresponding to each pixel in the target remote sensing image belongs to each land use type.

[0009] The land segmentation model is trained based on all land use classification probability maps to obtain a trained land segmentation model;

[0010] The trained land segmentation model is used to segment the remote sensing image to be classified, and the land use classification results of the remote sensing image to be classified are obtained.

[0011] Optionally, the land use segmentation model includes a patch embedding module, a spatial encoding module, a first encoding module, a second encoding module, a third encoding module, a first continuous low-frequency feature decomposition module, a second continuous low-frequency feature decomposition module, a third continuous low-frequency feature decomposition module, a first decoding module, a second decoding module, a third decoding module, and a semantic segmentation module;

[0012] The input of the patch embedding module is the input of the land use segmentation model, and the output of the semantic segmentation module is the output of the land use segmentation model;

[0013] The output end of the patch embedding module is respectively connected to the input end of the spatial encoding module and the first input end of the first continuous low-frequency feature decomposition module; the output end of the spatial encoding module is respectively connected to the first input end of the first encoding module and the input end of the first decoding module; the first output end of the first continuous low-frequency feature decomposition module is connected to the second input end of the first encoding module; the second output end of the first continuous low-frequency feature decomposition module is connected to the input end of the second decoding module; the third output end of the first continuous low-frequency feature decomposition module is connected to the first input end of the second continuous low-frequency feature decomposition module; the output end of the first encoding module is respectively connected to the first input end of the second encoding module and the second input end of the first continuous low-frequency feature decomposition module; the first output end of the second continuous low-frequency feature decomposition module is connected to the second input end of the second encoding module; the second continuous low-frequency feature decomposition module is connected to the second input end of the second encoding module. The second output end of the module is connected to the input end of the third decoding module, the third output end of the second continuous low-frequency feature decomposition module is connected to the first input end of the third continuous low-frequency feature decomposition module, the output end of the second encoding module is respectively connected to the first input end of the third encoding module and the second input end of the second continuous low-frequency feature decomposition module, the first output end of the third continuous low-frequency feature decomposition module is connected to the second input end of the third encoding module, the output end of the third decoding module is connected to the second input end of the third continuous low-frequency feature decomposition module, the third output end of the third continuous low-frequency feature decomposition module is connected to the input end of the third decoding module, the output end of the third decoding module is connected to the input end of the second decoding module, the output end of the second decoding module is connected to the input end of the first decoding module, and the output end of the first decoding module is connected to the input end of the semantic segmentation module.

[0014] Optionally, the patch embedding module includes a first convolutional layer, a first normalization layer, a GELU activation function layer, a second convolutional layer, and a second normalization layer connected in sequence;

[0015] The input of the first convolutional layer is the input of the patch embedding module, and the output of the second normalization layer is the output of the patch embedding layer.

[0016] Optionally, the first encoding module, the second encoding module, and the third encoding module are all encoding modules;

[0017] The coding module includes a patch merging layer, a splicing layer, and a spatial and frequency fusion coding layer connected in sequence;

[0018] The input end of the patch merging layer is the first input end of the encoding module, the input end of the splicing layer is the second input end of the encoding module, and the spatial and frequency fusion encoding layer is the output end of the encoding module.

[0019] Optionally, the first continuous low-frequency feature decomposition module, the second continuous low-frequency feature decomposition module, and the third continuous low-frequency feature decomposition module are all continuous low-frequency feature decomposition modules;

[0020] The continuous low-frequency feature decomposition module includes a discrete wavelet transform layer, a first normalization layer, a second normalization layer, a global-local attention layer, a channel attention layer, a convolution layer, and a feature concatenation layer;

[0021] The input end of the discrete wavelet transform layer is the first input end of the continuous low-frequency feature decomposition module, the input end of the feature splicing layer is the second input end of the continuous low-frequency feature decomposition module, the output end of the global-local attention module is the first output end of the continuous low-frequency feature decomposition module, the output end of the feature splicing layer is the second output end of the continuous low-frequency feature decomposition module, and the output end of the first normalization layer is the third output end of the continuous low-frequency feature decomposition module;

[0022] The first output of the discrete wavelet transform layer is connected to the input of the first normalization layer, the output of the first normalization layer is connected to the input of the global local attention layer, the second output of the discrete wavelet transform layer is connected to the input of the second normalization layer, the output of the second normalization layer is connected to the input of the channel attention layer, the output of the channel attention layer is connected to the input of the convolution layer, and the output of the convolution layer is connected to the input of the feature splicing layer.

[0023] Optionally, the semantic segmentation module includes a first semantic convolution layer, a first bilinear interpolation layer, a second semantic convolution layer, a second bilinear interpolation layer, and a segmentation convolution layer connected in sequence;

[0024] The input end of the first semantic convolution layer is the input end of the semantic segmentation module, and the output end of the segmentation convolution layer is the output end of the semantic segmentation module.

[0025] Optionally, a land segmentation model is trained based on all land use classification probability maps to obtain a trained land segmentation model, including:

[0026] Construct a loss function that describes the accuracy of all land use classification probability maps based on all land use classification probability maps;

[0027] Determine whether the value of the loss function is less than the preset value of the loss function;

[0028] If so, the land segmentation model is used as the trained land segmentation model;

[0029] Otherwise, the land segmentation model is adjusted, and the step of continuously decomposing low-frequency features and semantic segmentation of each target remote sensing image using the land use segmentation model is returned to obtain a land use classification probability map of each target remote sensing image.

[0030] In a second aspect, an embodiment of the present application provides a land use classification device based on low-frequency features, comprising:

[0031] A first acquisition module is used to acquire multiple target remote sensing images;

[0032] The semantic segmentation module is used to obtain the low-frequency features of each target remote sensing image using the land use segmentation model, and continuously decompose and decode each low-frequency feature to obtain the final decoded features of each target remote sensing image. Each final decoded feature is semantically segmented to obtain a land use classification probability map for each target remote sensing image; the land use classification probability map is used to describe the probability that the area corresponding to each pixel in the target remote sensing image belongs to each land use type;

[0033] A training module is used to train the land segmentation model based on all land use classification probability maps to obtain a trained land segmentation model;

[0034] A segmentation module is used to segment the remote sensing image to be classified using the trained land segmentation model to obtain a land use classification probability map of the remote sensing image to be classified;

[0035] The second acquisition module is used to obtain the land use classification result of the remote sensing image to be classified based on the land use classification probability map of the remote sensing image to be classified; the land use classification result is used to describe the land use status of the area corresponding to the remote sensing image to be classified.

[0036] In a third aspect, an embodiment of the present application provides a terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned land use classification method based on low-frequency features when executing the above-mentioned computer program.

[0037] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned land use classification method based on low-frequency features.

[0038] The above solution of the present application has the following beneficial effects:

[0039] In an embodiment of the present application, a plurality of target remote sensing images are acquired, and then a land use segmentation model is used to acquire the low-frequency features of each target remote sensing image, and each low-frequency feature is continuously decomposed and decoded to obtain the final decoded features of each target remote sensing image, and each final decoded feature is semantically segmented to obtain a land use classification probability map of each target remote sensing image, and then the land segmentation model is trained based on all land use classification probability maps to obtain a trained land segmentation model, and finally the trained land segmentation model is used to segment the remote sensing image to be classified to obtain the land use classification result of the remote sensing image to be classified. Among them, the continuous decomposition of the low-frequency features of the target remote sensing image by the land use segmentation model can enhance the characterization capability of the low-frequency features of the target remote sensing image for complex land objects, provide rich feature information for land use classification, and thus improve the accuracy of land use classification, and the training of the land use segmentation model can improve the performance of the land use segmentation model, and the land use classification of the remote sensing image to be classified by using the trained land use segmentation model can further improve the accuracy of land use classification.

[0040] Other beneficial effects of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0042] Figure 1 A flow chart of a land use classification method based on low-frequency features provided in one embodiment of the present application;

[0043] Figure 2 A schematic diagram of the structure of a land use segmentation model provided in one embodiment of the present application;

[0044] Figure 3 A schematic diagram of the structure of the encoding module provided in one embodiment of the present application;

[0045] Figure 4 A schematic diagram of the structure of a continuous low-frequency feature decomposition module provided in one embodiment of the present application;

[0046] Figure 5 A schematic diagram of the structure of a land use classification device based on low-frequency features provided in one embodiment of the present application;

[0047] Figure 6 A schematic diagram of the structure of a terminal device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0048] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0049] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0050] It will also be understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0051] As used in this specification and the appended claims, the term "if" can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0052] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0053] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0054] In response to the problem of low accuracy of existing land use classification, an embodiment of the present application provides a land use classification method based on low-frequency features. The land use classification method obtains the low-frequency features of each target remote sensing image by using a land use segmentation model, and continuously decomposes and decodes each low-frequency feature to obtain the final decoded features of each target remote sensing image, performs semantic segmentation on each final decoded feature, obtains a land use classification probability map of each target remote sensing image, and then trains the land segmentation model based on all land use classification probability maps to obtain a trained land segmentation model. Finally, the trained land segmentation model is used to segment the remote sensing image to be classified to obtain the land use classification result of the remote sensing image to be classified. Among them, the continuous decomposition of the low-frequency features of the target remote sensing image by the land use segmentation model can enhance the low-frequency features of the target remote sensing image to represent complex land objects, provide rich feature information for land use classification, and thus improve the accuracy of land use classification. Training the land use segmentation model can improve the performance of the land use segmentation model. Using the trained land use segmentation model to classify the remote sensing image to be classified further improves the accuracy of land use classification.

[0055] Next, the land use classification method based on low-frequency features provided in this application is exemplified.

[0056] like Figure 1 As shown, the land use classification method based on low-frequency features provided in this application includes the following steps:

[0057] Step 11: Acquire multiple target remote sensing images.

[0058] Multiple target remote sensing images correspond one-to-one to multiple regions, and a region can be a village, an urban area, etc.

[0059] For example, a target remote sensing image can be obtained by accessing a website that publicly provides remote sensing images. Each target remote sensing image has corresponding label data that describes the actual land use category to which the area corresponding to each pixel in the target remote sensing image belongs. For example, the label data is stored in the form of a raster file and mapped to a specific category based on a color mapping table of the label data, thereby generating single-band label data, where a value of 0 represents background, a value of 1 represents category 1 (e.g., one of rice fields, irrigated land, garden land, and dry land), a value of 2 represents category 2 (e.g., another of rice fields, irrigated land, garden land, and dry land that is different from category 1), and so on.

[0060] Step 12: Use the land use segmentation model to obtain the low-frequency features of each target remote sensing image, and continuously decompose and decode each low-frequency feature to obtain the final decoded features of each target remote sensing image. Perform semantic segmentation on each final decoded feature to obtain a land use classification probability map of each target remote sensing image.

[0061] The above land use classification probability map is used to describe the probability that the area corresponding to each pixel in the target remote sensing image belongs to each land use type.

[0062] like Figure 2 As shown, the land use segmentation model includes a patch embedding module, a spatial encoding module, a first encoding module, a second encoding module, a third encoding module, a first continuous low-frequency feature decomposition module, a second continuous low-frequency feature decomposition module, a third continuous low-frequency feature decomposition module, a first decoding module, a second decoding module, a third decoding module, and a semantic segmentation module.

[0063] The input of the patch embedding module is the input of the land use segmentation model, and the output of the semantic segmentation module is the output of the land use segmentation model.

[0064] The output end of the patch embedding module is respectively connected to the input end of the spatial encoding module and the first input end of the first continuous low-frequency feature decomposition module; the output end of the spatial encoding module is respectively connected to the first input end of the first encoding module and the input end of the first decoding module; the first output end of the first continuous low-frequency feature decomposition module is connected to the second input end of the first encoding module; the second output end of the first continuous low-frequency feature decomposition module is connected to the input end of the second decoding module; the third output end of the first continuous low-frequency feature decomposition module is connected to the first input end of the second continuous low-frequency feature decomposition module; the output end of the first encoding module is respectively connected to the first input end of the second encoding module and the second input end of the first continuous low-frequency feature decomposition module; the first output end of the second continuous low-frequency feature decomposition module is connected to the second input end of the second encoding module; the second continuous low-frequency feature decomposition module is connected to the second input end of the second encoding module. The second output end of the module is connected to the input end of the third decoding module, the third output end of the second continuous low-frequency feature decomposition module is connected to the first input end of the third continuous low-frequency feature decomposition module, the output end of the second encoding module is respectively connected to the first input end of the third encoding module and the second input end of the second continuous low-frequency feature decomposition module, the first output end of the third continuous low-frequency feature decomposition module is connected to the second input end of the third encoding module, the output end of the third decoding module is connected to the second input end of the third continuous low-frequency feature decomposition module, the third output end of the third continuous low-frequency feature decomposition module is connected to the input end of the third decoding module, the output end of the third decoding module is connected to the input end of the second decoding module, the output end of the second decoding module is connected to the input end of the first decoding module, and the output end of the first decoding module is connected to the input end of the semantic segmentation module.

[0065] It should be noted that the patch embedding module is used to perform patch embedding on the input data, the spatial encoding module is used to encode the spatial dimension of the input data, the first encoding module, the second encoding module, and the third encoding module are all used to fuse and encode all input data, the first continuous low-frequency feature decomposition module, the second continuous low-frequency feature decomposition module, and the third continuous low-frequency feature decomposition module are used to perform low-frequency feature decomposition on the target remote sensing image processed by the patch embedding module, and the target remote sensing image processed by the patch embedding module is used as the input data of the first continuous low-frequency feature decomposition module. After the low-frequency feature decomposition is performed in the first continuous low-frequency feature decomposition module, the downstream second continuous low-frequency feature decomposition module performs low-frequency feature decomposition on the output data of the first continuous low-frequency feature decomposition module again, and the third continuous low-frequency feature decomposition module downstream of the second continuous low-frequency feature decomposition module performs low-frequency feature decomposition on the output data of the second continuous low-frequency feature decomposition module again. The first decoding module, the second decoding module, and the third decoding module are all used to decode all input data, and the output data of the first decoding module is the final decoding feature. The semantic segmentation module is used to perform semantic segmentation of land use types according to the input data and output a land use classification probability map.

[0066] The patch embedding module includes a first convolutional layer, a first normalization layer, a GELU activation function layer, a second convolutional layer, and a second normalization layer connected in sequence. The input end of the first convolutional layer is the input end of the patch embedding module, and the output end of the second normalization layer is the output end of the patch embedding module. The first convolutional layer and the second convolutional layer are both used to perform convolution operations on the input data, the first normalization layer and the second normalization layer are both used to normalize the input data, and the GELU activation function layer is used to perform GELU activation function operations on the input data. Exemplarily, the first convolutional layer has 3 input channels, 48 output channels, a filter kernel size of 3, a step size of 2, and a padding of 1. The second convolutional layer has 48 input channels, 96 output channels, a filter kernel size of 3, a step size of 2, and a padding of 1.

[0067] The spatial coding module includes a first spatial convolution layer and a second spatial convolution layer connected in sequence. The input end of the first spatial convolution layer is the input end of the spatial coding module, and the output end of the second spatial convolution layer is the output end of the spatial coding module. The first spatial convolution layer and the second spatial convolution layer are both convolution layers. The convolution layer includes a convolution block, a batch normalization block, and a Relu activation function block connected in sequence. The input end of the convolution block is the input end of the spatial convolution layer, and the output end of the Relu activation function block is the output end of the spatial convolution layer. The convolution block is used to perform convolution operations on the input data, the batch normalization block is used to normalize the input data, and the Relu activation function block is used to perform Relu activation function operations on the input data. Exemplarily, the filter size in the above convolution block is 3, the step size is 1, and the padding is 1.

[0068] The first encoding module, the second encoding module, and the third encoding module are all encoding modules.

[0069] like Figure 3 As shown, the coding module includes a patch merging layer, a splicing layer, and a spatial and frequency fusion coding layer connected in sequence.

[0070] The input end of the patch merging layer is the first input end of the encoding module, the input end of the splicing layer is the second input end of the encoding module, and the spatial and frequency fusion encoding layer is the output end of the encoding module.

[0071] It should be noted that the patch merging layer is used to perform convolution and normalization operations on the input data, including patch convolution blocks and layer normalization blocks connected in sequence. The input end of the patch convolution block is the input end of the patch merging layer, and the output end of the layer normalization block is the output end of the patch merging layer. The patch convolution block is used to perform convolution operations on the input data, and the layer normalization block is used to perform normalization operations on the input data. Exemplarily, the patch convolution block has 96 input channels, 192 output channels, a filter kernel size of 3, a stride of 2, and a padding of 1. The splicing layer is used to splice the input data, and the spatial and frequency coding layer is used to encode the input data in the spatial and frequency dimensions.

[0072] The spatial and frequency fusion coding layer includes a first coding convolution block and a second coding convolution block connected in sequence. The input end of the first coding convolution block is the input end of the spatial and frequency fusion coding layer, and the output end of the second coding convolution block is the output end of the spatial and frequency fusion coding layer. The first coding convolution block and the second coding convolution block are both coding convolution blocks. The coding convolution block includes a convolution unit, a batch normalization unit, and a Relu activation function unit connected in sequence. The input end of the convolution unit is the input end of the coding convolution block, and the output end of the Relu activation function unit is the output end of the coding convolution block. The convolution unit is used to perform convolution operations on the input data, the batch normalization unit is used to normalize the input data, and the Relu activation function unit is used to perform Relu activation function operations on the input data. Exemplarily, the filter size in the above-mentioned convolution unit is 3, the step size is 1, and the padding is 1.

[0073] The first continuous low-frequency feature decomposition module, the second continuous low-frequency feature decomposition module, and the third continuous low-frequency feature decomposition module are all continuous low-frequency feature decomposition modules.

[0074] like Figure 4 As shown in the figure, the continuous low-frequency feature decomposition module includes a discrete wavelet transform layer, a first normalization layer, a second normalization layer, a global-local attention layer, a channel attention layer, a convolution layer, and a feature splicing layer.

[0075] The input end of the discrete wavelet transform layer is the first input end of the continuous low-frequency feature decomposition module, the input end of the feature splicing layer is the second input end of the continuous low-frequency feature decomposition module, the output end of the global-local attention module is the first output end of the continuous low-frequency feature decomposition module, the output end of the feature splicing layer is the second output end of the continuous low-frequency feature decomposition module, and the output end of the first normalization layer is the third output end of the continuous low-frequency feature decomposition module.

[0076] The first output of the discrete wavelet transform layer is connected to the input of the first normalization layer, the output of the first normalization layer is connected to the input of the global local attention layer, the second output of the discrete wavelet transform layer is connected to the input of the second normalization layer, the output of the second normalization layer is connected to the input of the channel attention layer, the output of the channel attention layer is connected to the input of the convolution layer, and the output of the convolution layer is connected to the input of the feature splicing layer.

[0077] The output data of the third output terminal of the first continuous low-frequency feature decomposition module is the normalized low-frequency feature.

[0078] It should be noted that for the third continuous low-frequency feature decomposition module, the module structure is the same as the first continuous low-frequency feature decomposition module and the second continuous low-frequency feature decomposition module, but the third output end of the module is not connected, that is, the output data of the third output end is not used in the land segmentation model of this application.

[0079] The discrete wavelet transform layer is used to decompose the input data in the horizontal and vertical directions to obtain the feature information of the high-frequency and low-frequency parts. The first normalization layer and the second normalization layer are both used to normalize the input data. The feature information of the low-frequency part is input into the first normalization layer through the first output end of the discrete wavelet transform layer, and the feature information of the high-frequency part is input into the second normalization layer through the second output end of the discrete wavelet transform layer. The global-local attention layer is used to perform global attention mechanism and local attention mechanism operations on the input data. The channel attention layer is used to perform channel attention operations on the input data. The convolution layer is used to perform convolution operations on the input data. The feature splicing layer is used to splice all input data.

[0080] For example, taking the first continuous low-frequency feature decomposition module as an example, the discrete wavelet transform layer uses the Haar basis and decomposes in the horizontal and vertical directions. First, the input data F1 is decomposed in the horizontal direction. For any position F1(x,y) in F1, w represents the width of the target remote sensing image, and h represents the high and low frequency parts of the target remote sensing image:

[0081]

[0082] High frequency part:

[0083]

[0084] in and The dimensions are

[0085] Then decompose each column again to get the low-frequency part:

[0086]

[0087] Vertical high frequency part:

[0088]

[0089] Horizontal high frequency part:

[0090]

[0091] Diagonal high frequency part:

[0092]

[0093] After decomposition is completed, low-frequency information is obtained and high-frequency information and Its size is The low-frequency information Input into the first normalization layer to obtain normalized low-frequency information Then the normalized low-frequency information Input into the global local attention module to extract the ground feature category information in the low-frequency features to obtain low-frequency features The low-frequency information and low-frequency characteristics The size of Low-frequency characteristics It is input into the corresponding splicing layer and spliced with the output data of the corresponding patch merging layer to obtain the splicing feature

[0094] The second normalization layer converts high-frequency information and After splicing according to the channel dimension, the spliced high-frequency information is obtained Its size is High-frequency information Input into the channel attention module to filter redundant high-frequency information and obtain high-frequency features Its size is Then the high frequency features Input into the convolution layer to obtain redundant high-frequency features Its size is

[0095] Based on the above description of the continuous low-frequency feature decomposition module and the encoding module, the data transmission between the two is illustrated by taking the first continuous low-frequency feature decomposition module and the first encoding module as an example. In the first continuous low-frequency feature decomposition module, the discrete wavelet transform layer performs discrete wavelet transform on the input data to obtain high-frequency features and low-frequency features, wherein the low-frequency features are input to the first normalization layer for normalization operation, and the obtained output data reaches the input end of the second decoding module through the third output end. At the same time, the output data is input to the global-local attention layer for global and local attention operations. The obtained operation result reaches the splicing layer in the first encoding module through the first output end, and is spliced with the output data of the patch merging layer. The splicing result is input into the spatial and frequency fusion encoding layer. The output data of the spatial and frequency fusion encoding layer is input into the feature splicing layer of the first continuous low-frequency feature decomposition module, and is spliced with the high-frequency features processed by the second normalization layer, the channel attention layer, and the convolution layer. The splicing result is output to the second continuous low-frequency feature decomposition module through the second output end.

[0096] The first decoding module, the second decoding module, and the third decoding module are all decoding modules. The decoding modules include a nearest neighbor upsampling layer, a decoding convolution layer, a batch normalization layer, and a Relu activation function layer connected in sequence. The input end of the nearest neighbor upsampling layer is the input end of the decoding module, and the output end of the Relu activation function layer is the output end of the decoding module. The nearest neighbor upsampling layer is used to upsample the first input data (for the first decoding module, this data is the output data of the second decoding module, for the second decoding module, this data is the output data of the third decoding module, and for the third decoding module, this data is the output data of the third continuous low-frequency feature decomposition module) based on the nearest neighbor difference method, and concatenate the upsampled result with the other input data (for the first decoding module, this data is the output data of the spatial encoding module, for the second decoding module, this data is the output data of the first continuous low-frequency feature decomposition module, and for the third decoding module, this data is the output data of the second continuous low-frequency feature decomposition module). The decoding convolution layer is used to perform a convolution operation on the input data, the batch normalization layer is used to normalize the input data, and the Relu activation function layer is used to perform a Relu activation function operation on the input data. Exemplarily, the filter kernel size in the decoding convolution layer is 1 and the padding is 0.

[0097] The semantic segmentation module includes a first semantic convolution layer, a first bilinear interpolation layer, a second semantic convolution layer, a second bilinear interpolation layer, and a segmentation convolution layer connected in sequence. The input end of the first semantic convolution layer is the input end of the semantic segmentation module, and the output end of the segmentation convolution layer is the output end of the semantic segmentation module. The first semantic convolution layer and the second semantic convolution layer are both convolution layers (the same as the convolution layer structure in the spatial coding module). The first bilinear interpolation layer and the second bilinear interpolation layer are both used to perform bilinear interpolation operations on the input data, and the segmentation convolution layer is used to perform convolution operations on the input data. Exemplarily, the filter kernel size in the convolution layer is 3, the step size is 1, and the padding is 1. The filter kernel size of the convolution in the segmentation convolution layer is 7, the step size is 1, and the padding is 3. The output land use classification probability map FP has a size of C_N×w×h, where C_N is the number of land use classification categories.

[0098] It is worth mentioning that the continuous low-frequency feature decomposition module can convert spatial information into frequency domain information and decompose the low-frequency information therein multiple times. It provides the model with a more differentiated expression of spatial information and improves the recognition ability of different land object categories. The continuous low-frequency feature decomposition module is flexibly designed and can be seamlessly integrated with existing methods. Through global and local attention mechanisms, it provides the encoding module with low-frequency information representing a stable structure; at the same time, it uses channel attention and convolutional layers to remove redundant high-frequency features and provide the decoding module with frequency domain information describing image details (such as shape). It achieves a deep fusion of frequency domain features and spatial features, further enhancing the land segmentation model's perception of different land use types.

[0099] Step 13: Training the land segmentation model based on all land use classification probability maps to obtain a trained land segmentation model.

[0100] Specifically, a loss function describing the accuracy of all land use classification probability maps is constructed based on all land use classification probability maps.

[0101] Determine whether the value of the loss function is less than the preset value of the loss function.

[0102] If so, the land segmentation model is used as the trained land segmentation model.

[0103] Otherwise, the land segmentation model is adjusted, and the step of continuously decomposing low-frequency features and semantic segmentation of each target remote sensing image using the land use segmentation model is returned to obtain a land use classification probability map of each target remote sensing image.

[0104] It should be noted that the above loss function can be a multi-class cross entropy loss function constructed based on the label data of all land use classification probability maps and all target remote sensing images. When adjusting the land segmentation model, the parameters in the land segmentation model (such as the convolution kernel size, step size, etc.) can be adjusted. The land segmentation model can also be trained based on the loss function using the AdamW optimizer and a cosine annealing learning rate. For example, the initial learning rate is set to 0.001, the batch size is set to 8, and the cycle is set to 128. The cycle length of the cosine annealing learning rate is set to 16.

[0105] Step 14: Segment the remote sensing image to be classified using the trained land segmentation model to obtain a land use classification result of the remote sensing image to be classified.

[0106] The remote sensing image to be classified is a remote sensing image of the area requiring land use classification. The land use classification result is used to describe the land use type of the area corresponding to the remote sensing image to be classified. For example, the land use classification result is a label for each pixel, which is used to describe the land use type of the land area corresponding to the pixel.

[0107] Specifically, the remote sensing image to be classified is input into the trained land segmentation model, and the trained land use segmentation model is used to obtain the low-frequency features of the remote sensing image to be classified, and the low-frequency features are continuously decomposed and decoded to obtain the final decoded features of the remote sensing image to be classified. The final decoded features are semantically segmented to obtain the land use classification probability map of the remote sensing image to be classified. Finally, the torch.argmax() function can be used to calculate the land use classification probability map to obtain the land use classification result.

[0108] It is worth mentioning that the use of land use segmentation model to continuously decompose the low-frequency features of the target remote sensing image can enhance the characterization ability of the low-frequency features of the target remote sensing image for complex landforms, provide rich feature information for land use classification, and thus improve the accuracy of land use classification. Training the land use segmentation model can improve the performance of the land use segmentation model. Using the trained land use segmentation model to classify the remote sensing images to be classified can further improve the accuracy of land use classification.

[0109] Furthermore, the land use segmentation model extracts contextual information from remote sensing images by continuously decomposing low-frequency features, promoting the fusion of spatial and frequency domain features, thereby improving the accuracy of land use classification. This method utilizes continuous low-frequency features to achieve a more differentiated representation from spatial information by continuously decomposing low-frequency information, thus enhancing the model's ability to perceive different land feature types.

[0110] The method of the present application is illustrated below with reference to a specific example.

[0111] We tested the GID dataset, a large-scale, high-resolution remote sensing land cover dataset based on data from my country's Gaofen-2 satellite. The GID dataset consists of two parts: a large-scale classification set (GID-5) and a fine-grained land cover set (GID-15). The fine-grained land cover set (GID-15) contains 15 categories, including rice paddies, irrigated land, dry land, gardens, arbor forests, shrub forests, natural grasslands, artificial grasslands, industrial land, urban residential areas, rural residential areas, transportation land, rivers, lakes, and ponds, totaling 30,000 image patches.

[0112] Taking RS-Mamba, a remote sensing image classification model based on a state-space model, as the benchmark model, the continuous low-frequency feature decomposition module proposed in this application is embedded into it, providing the encoding module with low-frequency information representing the stationary structure, and providing the decoding module with redundant high-frequency information representing image details (such as shape). This method achieves a deep fusion of spatial and frequency domain features. In the GID-15 dataset, the method of this application achieved an average F1 score (Score) of 87.69% and an average intersection over union (mIoU) of 79.16%. Compared with the baseline model, comprehensive improvements were achieved in all 15 categories, with the average IoU increased by 2.09%. In particular, the intersection over union (IoU) of gardens, artificial grasslands and transportation land categories increased by 3.97%, 3.4% and 4.44% respectively. The test results are shown in Table 1.

[0113]

[0114]

[0115] Table 1

[0116] Here, Ours represents the method of this application.

[0117] This shows that the land use type classification using the method of the present application is highly accurate.

[0118] The following is an exemplary description of the land use classification device based on low-frequency features provided in this application.

[0119] like Figure 4 As shown, an embodiment of the present application provides a land use classification device based on low-frequency features. The land use classification device based on low-frequency features 400 includes:

[0120] The first acquisition module 401 is used to acquire multiple target remote sensing images;

[0121] Semantic segmentation module 402 is used to obtain low-frequency features of each target remote sensing image using a land use segmentation model, and continuously decompose and decode each low-frequency feature to obtain a final decoded feature of each target remote sensing image. Semantic segmentation is performed on each final decoded feature to obtain a land use classification probability map for each target remote sensing image. The land use classification probability map is used to describe the probability that the area corresponding to each pixel in the target remote sensing image belongs to each land use type.

[0122] A training module 403 is used to train the land segmentation model based on all land use classification probability maps to obtain a trained land segmentation model;

[0123] The segmentation module 404 is used to segment the remote sensing image to be classified using the trained land segmentation model to obtain a land use classification result of the remote sensing image to be classified.

[0124] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of this application. Their specific functions and technical effects can be found in the method embodiment section and will not be repeated here.

[0125] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0126] like Figure 5 As shown, an embodiment of the present application provides a terminal device, and the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 5Only one processor is shown in the figure), a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 implements the steps of any of the above method embodiments when executing the computer program D102.

[0127] Specifically, when the processor D100 executes the computer program D102, it obtains multiple target remote sensing images, then uses the land use segmentation model to obtain the low-frequency features of each target remote sensing image, and continuously decomposes and decodes each low-frequency feature to obtain the final decoded features of each target remote sensing image, performs semantic segmentation on each final decoded feature, obtains a land use classification probability map of each target remote sensing image, and then trains the land segmentation model based on all land use classification probability maps to obtain a trained land segmentation model. Finally, the trained land segmentation model is used to segment the remote sensing image to be classified to obtain the land use classification result of the remote sensing image to be classified. Among them, using the land use segmentation model to continuously decompose the low-frequency features of the target remote sensing image can enhance the low-frequency features of the target remote sensing image to represent complex land objects, provide rich feature information for land use classification, and thus improve the accuracy of land use classification. Training the land use segmentation model can improve the performance of the land use segmentation model. Using the trained land use segmentation model to classify the remote sensing image to be classified further improves the accuracy of land use classification.

[0128] The processor D100 may be a central processing unit (CPU), or may be another general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.

[0129] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may also be an external storage device of the terminal device D10, such as a plug-in hard disk, a smart memory card (SMC, SmartMedia Card), a secure digital (SD, Secure Digital) card, a flash card, etc. equipped on the terminal device D10. Furthermore, the memory D101 may also include both an internal storage unit of the terminal device D10 and an external storage device. The memory D101 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program. The memory D101 may also be used to temporarily store data that has been output or is to be output.

[0130] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned various method embodiments can be implemented.

[0131] An embodiment of the present application provides a computer program product. When the computer program product is run on a terminal device, the terminal device can implement the steps in the above-mentioned method embodiments when executing the computer program product.

[0132] If the integrated unit is implemented as a software functional unit and sold or used as a standalone product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to the low-frequency feature-based land use classification method apparatus / terminal device, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium. Examples include USB flash drives, removable hard drives, magnetic disks, or optical disks. In some jurisdictions, based on legislation and patent practice, computer-readable media cannot be electric carrier signals or telecommunication signals.

[0133] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0134] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0135] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A land use classification method based on low-frequency features, characterized in that: include: Acquire multiple target remote sensing images; A land use segmentation model is used to obtain low-frequency features of each target remote sensing image, and each low-frequency feature is continuously decomposed and decoded to obtain final decoded features of each target remote sensing image. Each final decoded feature is semantically segmented to obtain a land use classification probability map for each target remote sensing image; the land use classification probability map is used to describe the probability that the area corresponding to each pixel in the target remote sensing image belongs to each land use type; Training the land segmentation model based on all land use classification probability maps to obtain a trained land segmentation model; The trained land segmentation model is used to segment the remote sensing image to be classified to obtain a land use classification result of the remote sensing image to be classified.

2. The land use classification method according to claim 1, characterized in that: The land use segmentation model includes a patch embedding module, a spatial encoding module, a first encoding module, a second encoding module, a third encoding module, a first continuous low-frequency feature decomposition module, a second continuous low-frequency feature decomposition module, a third continuous low-frequency feature decomposition module, a first decoding module, a second decoding module, a third decoding module, and a semantic segmentation module; The input end of the patch embedding module is the input end of the land use segmentation model, and the output end of the semantic segmentation module is the output end of the land use segmentation model; The output end of the patch embedding module is respectively connected to the input end of the spatial encoding module and the first input end of the first continuous low-frequency feature decomposition module; the output end of the spatial encoding module is respectively connected to the first input end of the first encoding module and the input end of the first decoding module; the first output end of the first continuous low-frequency feature decomposition module is connected to the second input end of the first encoding module; the second output end of the first continuous low-frequency feature decomposition module is connected to the input end of the second decoding module; the third output end of the first continuous low-frequency feature decomposition module is connected to the first input end of the second continuous low-frequency feature decomposition module; the output end of the first encoding module is respectively connected to the first input end of the second encoding module and the second input end of the first continuous low-frequency feature decomposition module; the first output end of the second continuous low-frequency feature decomposition module is connected to the second input end of the second encoding module; the second continuous low-frequency feature decomposition module The second output end of the module is connected to the input end of the third decoding module, the third output end of the second continuous low-frequency feature decomposition module is connected to the first input end of the third continuous low-frequency feature decomposition module, the output end of the second encoding module is respectively connected to the first input end of the third encoding module and the second input end of the second continuous low-frequency feature decomposition module, the first output end of the third continuous low-frequency feature decomposition module is connected to the second input end of the third encoding module, the output end of the third decoding module is connected to the second input end of the third continuous low-frequency feature decomposition module, the third output end of the third continuous low-frequency feature decomposition module is connected to the input end of the third decoding module, the output end of the third decoding module is connected to the input end of the second decoding module, the output end of the second decoding module is connected to the input end of the first decoding module, and the output end of the first decoding module is connected to the input end of the semantic segmentation module.

3. The land use classification method according to claim 2, characterized in that: The patch embedding module includes a first convolutional layer, a first normalization layer, a GELU activation function layer, a second convolutional layer and a second normalization layer connected in sequence; The input end of the first convolutional layer is the input end of the patch embedding module, and the output end of the second normalization layer is the output end of the patch embedding layer.

4. The land use classification method according to claim 2, wherein: The first encoding module, the second encoding module, and the third encoding module are all encoding modules; The coding module includes a patch merging layer, a splicing layer, and a spatial and frequency fusion coding layer connected in sequence; The input end of the patch merging layer is the first input end of the encoding module, the input end of the splicing layer is the second input end of the encoding module, and the spatial and frequency fusion encoding layer is the output end of the encoding module.

5. The land use classification method according to claim 2, characterized in that: The first continuous low-frequency feature decomposition module, the second continuous low-frequency feature decomposition module, and the third continuous low-frequency feature decomposition module are all continuous low-frequency feature decomposition modules; The continuous low-frequency feature decomposition module includes a discrete wavelet transform layer, a first normalization layer, a second normalization layer, a global-local attention layer, a channel attention layer, a convolution layer, and a feature splicing layer; The input end of the discrete wavelet transform layer is the first input end of the continuous low-frequency feature decomposition module, the input end of the feature splicing layer is the second input end of the continuous low-frequency feature decomposition module, the output end of the global-local attention module is the first output end of the continuous low-frequency feature decomposition module, the output end of the feature splicing layer is the second output end of the continuous low-frequency feature decomposition module, and the output end of the first normalization layer is the third output end of the continuous low-frequency feature decomposition module; The first output end of the discrete wavelet transform layer is connected to the input end of the first normalization layer, the output end of the first normalization layer is connected to the input end of the global local attention layer, the second output end of the discrete wavelet transform layer is connected to the input end of the second normalization layer, the output end of the second normalization layer is connected to the input end of the channel attention layer, the output end of the channel attention layer is connected to the input end of the convolution layer, and the output end of the convolution layer is connected to the input end of the feature splicing layer.

6. The land use classification method according to claim 2, characterized in that: The semantic segmentation module includes a first semantic convolution layer, a first bilinear interpolation layer, a second semantic convolution layer, a second bilinear interpolation layer, and a segmentation convolution layer connected in sequence; The input end of the first semantic convolution layer is the input end of the semantic segmentation module, and the output end of the segmentation convolution layer is the output end of the semantic segmentation module.

7. The land use classification method according to claim 1, characterized in that: The land segmentation model is trained based on all land use classification probability maps to obtain a trained land segmentation model, including: Construct a loss function that describes the accuracy of all land use classification probability maps based on all land use classification probability maps; Determine whether the value of the loss function is less than a preset value of the loss function; If yes, the land segmentation model is used as the trained land segmentation model; Otherwise, the land segmentation model is adjusted, and the process returns to the step of continuously decomposing low-frequency features and semantic segmenting each target remote sensing image using the land use segmentation model to obtain a land use classification probability map for each target remote sensing image.

8. A land use classification device based on low-frequency features, characterized in that: include: A first acquisition module is used to acquire multiple target remote sensing images; a semantic segmentation module for obtaining low-frequency features of each target remote sensing image using a land use segmentation model, continuously decomposing and decoding each low-frequency feature to obtain a final decoded feature of each target remote sensing image, and performing semantic segmentation on each final decoded feature to obtain a land use classification probability map for each target remote sensing image; the land use classification probability map is used to describe the probability that the area corresponding to each pixel in the target remote sensing image belongs to each land use type; A training module, configured to train the land segmentation model based on all land use classification probability maps to obtain a trained land segmentation model; A segmentation module is used to segment the remote sensing image to be classified using the trained land segmentation model to obtain a land use classification probability map of the remote sensing image to be classified; The second acquisition module is used to obtain the land use classification result of the remote sensing image to be classified based on the land use classification probability map of the remote sensing image to be classified; the land use classification result is used to describe the land use status of the area corresponding to the remote sensing image to be classified.

9. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the land use classification method based on low-frequency features according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the land use classification method based on low-frequency features according to any one of claims 1 to 7 is implemented.

Images