Multi-scale cultivated land change detection system based on frequency domain conversion

Abstract

The invention provides a multi-scale cultivated land change detection system based on frequency domain conversion, and belongs to the technical field of remote sensing image processing and computer vision. Firstly, feature extraction is carried out on a dual-time-phase image to obtain a multi-scale first feature map; performing multi-kernel parallel convolution processing on the first feature map through small convolution kernels with different sizes, and then splicing and fusing the first feature map into a second feature map; performing feature enhancement on the second feature map to generate an extraction feature map; and combining and classifying two paths of extracted feature maps corresponding to the double-time-phase image to obtain a detection map of cultivated land change. Meanwhile, the double-time-phase image for training is subjected to low-frequency feature replacement in the frequency domain, visual feature differences caused by different imaging conditions are effectively eliminated, and false change false detection is reduced. According to the method, false change interference is reduced through low-frequency component replacement, meanwhile, background noise caused by a large convolution kernel and information loss caused by cavity convolution can be avoided through multi-kernel parallel convolution processing of the small convolution kernel, and meanwhile the detection capacity of scale sensitive targets is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image processing and computer vision, and in particular relates to a multi-scale farmland change detection system based on frequency domain conversion. Background Art

[0002] The existing research on change detection of high-resolution remote sensing images has the following shortcomings:

[0003] First, the domain gap problem has not been effectively solved in high-resolution remote sensing image change detection. Existing technologies are highly sensitive to differences in imaging conditions, such as changes in illumination, perspective, and background complexity, resulting in a high false detection rate of pseudo changes. Traditional preprocessing and normalization methods, such as histogram matching, radiation correction, and feature normalization, lack adaptability and stability in complex scenes, and it is difficult to effectively eliminate the negative impact of changes in imaging conditions. In addition, deep learning models still have significant limitations in domain generalization capabilities. They usually rely on a large amount of labeled data for training, but their adaptability under different imaging conditions is limited, resulting in reduced performance in practical applications. Although adversarial generative networks and transfer learning methods have been explored in domain adaptation, their results are still not ideal and are difficult to be widely applied to changing imaging environments.

[0004] Secondly, in terms of multi-scale feature extraction, the large convolution kernels relied on by existing technologies are prone to introduce a large amount of irrelevant background information, interfering with the effective discrimination of the target area, and often leading to blurring of edge information and loss of details. This phenomenon is particularly evident in high-resolution remote sensing images, because high-resolution data itself contains rich detail information, and any blur or information loss will directly affect the accuracy of the detection results. At the same time, although dilated convolution can effectively expand the receptive field and capture a wider range of contextual information, it is easy to miss key local details due to the sparsity of its sampling method and structural damage, which in turn affects the accuracy and robustness of the detection. At the same time, the feature extraction method of a single scale is difficult to fully capture contextual information at different levels, and cannot achieve an effective balance between local and global relationships, resulting in limited performance of the model when processing complex scenes. Therefore, the existing multi-scale feature extraction schemes face the challenges of high computational cost, complex design and limited effect in practical applications. There is an urgent need for more efficient, accurate and adaptively adjustable feature extraction methods to improve the overall performance of high-resolution remote sensing image change detection.

[0005] In addition, existing technologies have significant deficiencies in capturing long-distance dependencies in changes in high-resolution remote sensing images. Local feature extraction methods, such as traditional convolutional neural networks (CNNs), are difficult to effectively integrate global context information, especially when dealing with large-scale changes. This is because the characteristics of local operations limit the model's ability to capture long-distance pixel relationships, making it difficult to ensure global consistency in complex scenes. Current attention mechanisms (such as self-attention mechanisms) mainly focus on enhancing local features and improving feature expression by weighting information in the neighborhood, but they are still insufficient in capturing long-distance dependencies across regions, and their high computational complexity and memory consumption make them difficult to be efficiently applied in large-scale high-resolution images. Although the traditional Siamese network has certain feature fusion capabilities and performs feature comparison through a dual-branch structure, its feature fusion method is relatively simple and lacks deep integration design. It cannot fully utilize multi-scale features and long-distance dependencies, which limits its application effect and robustness in high-resolution remote sensing image change detection tasks. Therefore, it is urgent to develop a new network architecture that can capture global context information and fine-grained features at the same time and has efficient computing capabilities to improve the overall performance and application value of high-resolution remote sensing image change detection.

[0006] In summary, the existing technologies have the following deficiencies in high-resolution remote sensing image change detection:

[0007] 1. Traditional methods are difficult to effectively deal with the differences in visual features caused by differences in imaging conditions (lighting, shadows, background, etc.) of dual-phase images, and are prone to false change misdetection.

[0008] 2. Existing multi-scale feature extraction methods such as dilated convolution and large convolution kernels are prone to losing key information or introducing background noise.

[0009] 3. Changing objects exhibit complex and diverse morphological representations, which are reflected in the different sizes and irregular shapes of changing ground objects, which makes it more challenging to simultaneously identify changing objects of different scales and accurately depict their changing boundaries.

[0010] These shortcomings limit the accuracy and robustness of the model in complex scenarios, making it difficult to be widely used in a variety of practical tasks, and increase the cost and difficulty of training and application. Summary of the invention

[0011] In view of this, the present invention proposes a multi-scale farmland change detection system based on frequency domain transformation, which aims to reduce pseudo-change interference, avoid background noise caused by large convolution kernels and information loss caused by void convolution, and improve the detection capability of scale-sensitive targets.

[0012] In order to solve the above technical problems, the present invention is implemented as follows.

[0013] A multi-scale farmland change detection system based on frequency domain conversion, the system includes a feature extraction model, a dual-path merging unit and a training unit;

[0014] The feature extraction model processes each channel of the input dual-phase image to obtain the sum of two extracted feature maps;

[0015] The feature extraction model includes a feature extraction module, a multi-core parallel convolution module, and a feature enhancement module; the feature extraction module is used to extract features from two images in the dual-phase image to obtain a multi-scale first feature map;

[0016] The multi-core parallel convolution module contains multiple parallel convolution branches. Each parallel convolution branch uses m convolution kernels of different sizes, and the convolution kernel is less than or equal to 13×13; the value range of m is 3-7; the first feature map enters each parallel convolution branch for processing, and then is spliced ​​and fused to obtain the second feature map ;

[0017] Feature enhancement module, used to enhance the second feature map Perform feature enhancement and generate extracted feature maps ;

[0018] The two-way merging unit extracts feature maps from two paths and Merge and classify to obtain the detection map P of cultivated land changes;

[0019] The training unit includes an FFT module and a multi-scale training module;

[0020] The FFT module is used to convert the dual-phase image from the spatial domain to the frequency domain when training the feature extraction model, unify the low-frequency areas in the dual-phase image by replacement, and then convert it back to the spatial domain to generate a dual-phase image after the low-frequency component is replaced as the dual-phase image used for training;

[0021] When the multi-scale training module trains the feature extraction model, the multi-scale first feature map of the feature extraction module is processed by the multi-core parallel convolution module, the feature enhancement module and the dual-channel merging unit to generate a multi-scale detection map P j , j represents the jth scale; using the detection map P of all scales j The loss is calculated jointly to optimize the feature extraction model. During inference detection, only the detection map P corresponding to the first feature map of the deepest layer is used as the inference result.

[0022] Preferably, when the FFT module replaces the low-frequency area of ​​the dual-phase image, for a series of dual-phase images, it randomly determines to copy the low-frequency area of ​​the first image in the dual-phase image to the second image, or to copy the low-frequency area of ​​the second image to the first image.

[0023] Preferably, the feature extraction module uses a ResNet-18 backbone network with the fully connected layer removed.

[0024] Preferably, the feature extraction module includes a first convolution layer Conv1 and four residual layers Res1, Res2, Res3 and Res4 connected in sequence; during training, first feature maps of three scales are extracted from the residual layers Res1, Res2 and Res4, and are respectively input into the multi-core parallel convolution module for processing to obtain second feature maps of three scales; during inference prediction, only the deepest first feature map output by the residual layer Res4 is used to input into the multi-core parallel convolution module for processing.

[0025] Preferably, in the feature extraction module, the first convolution layer Conv1 adopts a 7×7 convolution kernel, and the residual layers Res2 and Res3 are downsampled with a step size of 2.

[0026] Preferably, in the multi-core parallel convolution module, there are 5 parallel convolution branches, and the convolution kernels are 3×3, 5×5, 7×7, 9×9 and 11×11 respectively; the 5 processing results output by the parallel convolution branches are further fused through feature connection and convolution of the 1×1 convolution kernel to obtain a second feature map.

[0027] Preferably, the feature enhancement module enhances the feature representation of the changing area through position attention, spatial attention and cross attention mechanisms.

[0028] Preferably, the feature enhancement module adopts a cross-scale feature aggregator, which is composed of a position attention PA module, a spatial attention SA module, and an encoder and a decoder of a Transformer module;

[0029] The dimension of the PA module within the channel affects the second feature map Enhance the key positions and output the third feature map ; SA module performs multi-channel processing on the third feature map Enhance the key areas and generate feature map T; the encoder encodes the feature map T to generate a tag with context , passed to the decoder; the decoder uses a cross-attention mechanism to generate the third feature map of the PA module as the query, and the tokens from the encoder As key and value respectively, generate enhanced features, that is, extract feature maps .

[0030] Preferably, the PA module comprises a channel number uniform convolution layer, a feature adjustment module, a multi-channel point-by-point convolution layer, a normalization module and a weighting module connected in sequence;

[0031] The number of channels is unified in the convolution layer, which is used to convert the second feature map of the input into Unify to a fixed number of channels M, generate the first intermediate feature map X0 of M channels and output it to the feature adjustment module;

[0032] The feature adjustment module includes a batch normalization layer, a linear rectification activation function and a Dropout module, processes the first intermediate feature map X0, and generates a second intermediate feature map X1 of M channels;

[0033] The multi-channel point-by-point convolution layer includes M 1×1 convolution kernels, which perform convolution on corresponding channels of the second intermediate feature map X1 without sharing the convolution kernels, and generate the third intermediate feature X2;

[0034] The normalization module uses a Sigmoid activation function to normalize the value of each position of the third intermediate feature X2 to the range of [0,1], and generates a position attention weight map, in which each value represents the importance of the corresponding position in the first intermediate feature map X0;

[0035] The weighting module is used to perform element-by-element multiplication of the first intermediate feature map X0 and the position attention weight map to achieve position enhancement and obtain the third feature map .

[0036] Preferably, the two-way merging unit comprises a connection module and a classification module; the connection module extracts feature maps from the two paths and splicing; the classification module uses a convolutional layer to compare the spliced ​​feature maps to obtain a detection map of cultivated land changes.

[0037] Beneficial effects:

[0038] (1) Adaptive feature extraction and enhanced timeliness: Traditional change detection methods rely on technicians to manually select features. The algorithm effect is limited by experience and ability, and a large number of repeated experiments are required to determine parameters, making it difficult to ensure high robustness and timeliness. The system of the present invention effectively combines the powerful feature extraction and parameter optimization capabilities of deep convolutional neural networks, and can adaptively extract features based on data characteristics using iterative operations and gradient updates. In the application stage, feature extraction can be performed directly by loading the trained weight model without repeated experiments and manual tuning, which significantly improves timeliness and application value. Compared with traditional methods, the system of the present invention reduces reliance on manual intervention and avoids feature selection bias caused by insufficient experience of technicians. In addition, automated feature extraction and model optimization significantly shorten the development cycle and improve the speed and efficiency of model deployment.

[0039] (2) Multi-scale feature fusion optimization: Existing methods are prone to introducing background noise or losing key information when capturing multi-scale features, and it is difficult to take into account both fine-grained features and global modeling. The system of the present invention adopts a multi-core Inception module (MKI) to capture diverse feature information from local details to global structures through convolution kernels of different sizes, avoiding background noise caused by large convolution kernels and information loss caused by hole convolution. At the same time, the MKI module integrates multi-scale information, maintains the integrity and accuracy of features, and significantly improves the richness and accuracy of feature representation; traditional methods usually use convolution kernels of a single scale, which makes it difficult to fully capture multi-scale features in images, and easily leads to background noise interference and loss of important information. The system of the present invention not only enhances the feature expression capability through the design of parallel convolution branches, but also improves the understanding of complex scenes.

[0040] (3) Improved adaptability to complex scenes: Existing methods are difficult to cope with changes in visual features caused by differences in imaging conditions such as lighting, viewing angle and background, and are prone to false detection of pseudo changes. The system of the present invention introduces fast Fourier transform (FFT) to convert dual-phase images from the spatial domain to the frequency domain and replace the low-frequency components to unify the style of the two different phase images. This method effectively eliminates the differences in visual features caused by different imaging conditions and reduces false detection of pseudo changes. The random application of FFT enhances the generalization ability of the model, enabling it to adapt to more diverse data distributions, thereby performing more stably in practical applications. When processing images under different imaging conditions, traditional methods are prone to false detection of pseudo changes due to differences in visual features. The system of the present invention effectively eliminates these differences through frequency domain processing, allowing the model to maintain consistent detection performance under different imaging conditions, significantly improving robustness and reliability.

[0041] (4) Capturing long-distance dependencies: In a preferred embodiment, considering that existing methods are difficult to effectively capture long-distance dependencies in images, especially when dealing with large-scale changes, the feature enhancement module in the system of the present invention combines position attention (PA) and spatial attention (SA) to enhance the feature representation in key areas and focus on key areas of the entire image. The Transformer encoder (TE) and decoder (TD) are used to capture global context information and achieve effective understanding of long-distance dependencies. This design significantly improves the model's ability to understand and parse complex scenes. Traditional methods are often limited to local feature extraction and have difficulty capturing long-distance dependencies in images, which limits their performance in large-scale change detection. By introducing a multi-scale attention mechanism and a Transformer architecture, the system of the present invention can more effectively capture and utilize global context information, thereby significantly improving detection accuracy and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a schematic diagram of the multi-scale farmland change detection system based on frequency domain conversion of the present invention.

[0043] Figure 2 It is a block diagram of the composition of a multi-scale farmland change detection system based on frequency domain conversion in a preferred embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram of the FFT module.

[0045] Figure 4 Schematic diagram of multi-scale CNN and multi-core Inception modules.

[0046] Figure 5 Schematic diagram of skip connection of residual layer.

[0047] Figure 6 Schematic diagram of the cross-scale feature aggregator (CSFA).

[0048] Figure 7 Schematic diagram of the position attention (PA) module.

[0049] Figure 8 Schematic diagram of the Transformer encoder and decoder.

[0050] Fig. 9Visualization of experimental results on the CLCD dataset; white indicates true positive examples, black indicates true negative examples, purple indicates false positive examples, and green indicates false negative examples. (a) Phase 1; (b) Phase 2; (c) Label; (d) FC-EF; (e) FC-Siam-conc; (f) FC-Siam-diff; (g) DTCDSCN; (h) SNUNet; (i) BIT; (j) MSCANet; (k) SARASNet; (l) HATNet; (m) the proposed solution.

[0051] Fig.10 Visualization of experimental results on WHU-CD and LEVIR-CD datasets; white indicates true positives, black indicates true negatives, purple indicates false positives, and green indicates false negatives. (a) Phase 1; (b) Phase 2; (c) Label; (d) FC-EF; (e) FC-Siam-conc; (f) FC-Siam-diff; (g) DTCDSCN; (h) SNUNet; (i) BIT; (j) MSCANet; (k) SARASNet; (l) HATNet; (m) the proposed solution.

[0052] Fig.11 The figure is a visualization of the FFT effect; (a) the left side is the image of time stage 1, the middle is the image of time stage 2, and the right side is the image of time stage 1 after processing; (b) is the frequency domain image of the red band corresponding to (a).

[0053] Fig.12 Visualization of CSFA. (a) Bi-temporal image. (b) Multi-scale feature maps F1 and F2 extracted by MKI; (c) Attention maps P1 and P2 generated by PA; (d) Attention maps S1 and S2 generated by SA; (e) Fine feature maps T1 and T2 generated by Transformer. (f) Change map. DETAILED DESCRIPTION

[0054] The present invention provides a multi-scale farmland change detection system based on frequency domain conversion, and its core improvement is:

[0055] (1) During training, the present invention maps the dual-phase images from the spatial domain to the frequency domain through Fast Fourier Transform (FFT) and performs frequency domain feature interaction to optimize feature representation, thereby effectively alleviating the inevitable visual feature differences in change detection tasks and reducing pseudo-change interference.

[0056] (2) The present invention constructs a feature extraction module and a multi-core parallel convolution module, and uses an improved ResNet-18 and an Inception-style parallel small convolution kernel architecture to perform multi-scale feature extraction and progressive fusion, promote information coupling between features within the layer, and avoid background noise brought by large convolution kernels and information loss caused by void convolution.

[0057] (3) In a preferred embodiment, a cross-scale feature aggregator (CSFA) is designed to integrate multi-scale features and enhance contextual understanding through attention mechanisms of different scales and the Transformer structure, ensuring effective global modeling while capturing fine-grained features.

[0058] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0059] Figure 1 The schematic diagram of the multi-scale farmland change detection system based on frequency domain conversion of the present invention is shown as follows: Figure 1 As shown, the system includes a feature extraction model, a dual-path merging unit and a training unit.

[0060] The feature extraction model processes each channel of the input dual-phase image to obtain two extracted feature maps and ; The dual-path merging unit extracts feature maps from two paths and After merging and classification, we obtain the detection map P of cultivated land changes.

[0061] The feature extraction model includes a feature extraction module, a multi-core parallel convolution module, and a feature enhancement module. The feature extraction module is used to extract features from two images in the dual-phase image to obtain a multi-scale first feature map. , j represents the jth scale, i=1,2.

[0062] The multi-core parallel convolution module contains multiple parallel convolution branches, each of which uses m convolution kernels of different sizes, and the convolution kernel is less than or equal to 13×13; the value range of m is 3-7; the multi-scale first feature map Enter each parallel convolution branch for processing, and then perform splicing and fusion to obtain the second feature map .

[0063] Feature enhancement module, used to enhance the second feature map Perform feature enhancement and generate extracted feature maps .

[0064] The training unit includes an FFT module and a multi-scale training module; wherein,

[0065] The FFT module is used to convert the dual-phase image from the spatial domain to the frequency domain when training the feature extraction model, unify the low-frequency areas in the dual-phase image by replacement, and then convert it back to the spatial domain to generate a dual-phase image after the low-frequency component is replaced as the dual-phase image used for training;

[0066] When training the feature extraction model, the multi-scale training module extracts the multi-scale first feature map from multiple residual layers of the feature extraction module. , the first feature map of each scale After being processed by the multi-core parallel convolution module, feature enhancement module and dual-channel merging unit, a multi-scale detection map P is generated. j ; Using all scales of detection map P j The loss is calculated jointly to optimize the feature extraction model. During inference detection, only the detection map P corresponding to the first feature map of the deepest layer is used as the inference result.

[0067] In the preferred embodiment, the feature extraction module is implemented by a modified ResNet-18; the multi-core parallel convolution module is implemented by an Inception-style parallel small convolution kernel architecture; the feature extraction module and the multi-core parallel convolution module jointly realize multi-scale feature extraction and progressive fusion, alleviate the gradient vanishing problem, expand the receptive field while avoiding the introduction of background noise and information loss. These features jointly improve the accuracy and robustness of change detection, especially when processing complex high-resolution remote sensing images. The experimental results verify the effectiveness and superiority of the module, providing a solid foundation for future research and application.

[0068] The feature enhancement module uses a cross-scale feature aggregator (CSFA) to fuse the multi-scale features extracted in the early stage, and adaptively assigns different attention weights according to the scale difference of the target change, so as to effectively capture subtle spatial changes and contextual dependencies. CSFA can be applied in different computer vision tasks, such as image classification, target detection, semantic segmentation, etc.

[0069] The present invention further provides a preferred implementation of the above-mentioned multi-scale farmland change detection system, which specifically provides an architecture suitable for fine-grained feature extraction and global modeling under different imaging conditions - FIMANet, which effectively solves the problem of pseudo-change detection caused by differences in illumination, viewing angle and background in dual-phase images by combining fast Fourier transform (FFT), multi-scale feature extraction module (CNN), multi-core Inception module (MKI) and cross-scale feature aggregator (CSFA). This method converts the image from the spatial domain to the frequency domain through FFT, replaces the low-frequency components of the target image, and reduces the style differences caused by imaging conditions; the MKI module gradually extracts and refines multi-scale features to avoid background noise and information loss; the CSFA module captures global and local dependencies through attention mechanism and Transformer to improve detection accuracy. By mining the multi-scale features in change detection of high-resolution remote sensing images,

[0070] Figure 2 The functional block diagram of the preferred embodiment is shown below. Figure 2 The present invention is described in detail.

[0071] like Figure 2 As shown, the multi-scale farmland change detection system based on frequency domain conversion includes an FFT module, a multi-scale CNN module (Multi-Scale CNN), a multi-core Inception module (MKI, Multiple Kernel Inception), a cross-scale feature aggregator (CSFA, Cross-scale feature aggregator), and an output module (Output in the figure). The system should also include a multi-scale training module not shown in the figure. The system composed of these modules is called FIMANet (Frequency-Based Detailed Multi-Scale Approach for Remote Sensing Change Detection) in the present invention.

[0072] Among them, the FFT module and the multi-scale training module form a training unit, the multi-scale CNN module implements the function of the feature extraction module, the multi-core Inception module is a multi-core parallel convolution module, the CSFA implements the function of the feature enhancement module, and the output module implements the function of the dual-channel merging unit. The multi-scale CNN module, the multi-core Inception module, and the CSFA form a feature extraction model, in which the parameters and weights of each grid layer involved need to be trained, including but not limited to the convolution layer, batch normalization layer, Dropout layer behavior mentioned below, the weight of the Sigmoid activation function, the ResNet-18 backbone network, etc.

[0073] The training unit is responsible for model training. The FFT module is only used during training. During actual inference, the dual-phase image is directly input into the multi-scale CNN module. During training, the multi-scale CNN module extracts multi-scale feature maps. Each scale feature map is processed by the multi-core Inception module, CSFA and output module to obtain a multi-scale detection map. Then, the detection maps of all scales are used to generate the P j The loss is calculated jointly with the true value to optimize the feature extraction model. During actual reasoning, only the detection map P corresponding to the first feature map of the deepest layer is used as the reasoning result to obtain the predicted detection map P.

[0074] The following is a detailed description of each of the above modules.

[0075] (1) FFT module

[0076] To mitigate the domain differences between bi-temporal images caused by changes in imaging conditions (such as differences in illumination, shadows, and background), frequency domain feature interactions are used to reduce the domain gap and promote information coupling of intra-layer representations. This FFT module converts bi-temporal images from the spatial domain to the frequency domain, where regions with dramatic intensity changes correspond to high-frequency components (mapping the edges and contours of the image), while smoother regions correspond to low-frequency components (providing a comprehensive mapping of the overall image intensity). By replacing the frequency domain components of the bi-temporal in the frequency domain and then applying an inverse FFT to convert back to the spatial domain, the visual features of the two images can be effectively unified.

[0077] like Figure 3 As shown, the specific implementation process of the FFT module is as follows:

[0078] Input image preprocessing: Input dual-phase images, which are divided into two paths, respectively denoted as and , making sure they have been registered to ensure that pixels at the same geographic location are aligned.

[0079] Forward Fourier transform: for dual-phase images and 2D FFT (two-dimensional fast Fourier transform) is applied respectively to generate the corresponding spectrograms F1 and F2. The spectrogram shows the distribution of the image in the frequency domain, where the low-frequency components correspond to the overall brightness and smooth areas of the image, and the high-frequency components correspond to the edges and changes of the image. For example, the conversion formula for each channel i of the RGB three-channel image is as follows:

[0080] in and Represents the height and width of the image respectively, is the image space domain coordinate, is the image frequency domain coordinate, Representing images The corresponding spectrum The frequency value of the position, Representing images middle The pixel position.

[0081] Low-frequency component replacement: swap images in the frequency domain and The low-frequency component of one image is exchanged here to copy the low-frequency component of one image to another image. The low-frequency component is determined by a set threshold, and the threshold value is determined according to experimental results by enumerating on three public data sets. Let Δ represent the frequency domain sampling point size in the low-frequency area, and its value is determined by enumeration experiments. In one embodiment, the size of Δ is determined to be 10, see Table 2. Taking the replacement of the low-frequency component in the spectrum graph F1 with the corresponding low-frequency component in the spectrum graph F2 as an example, the formula is expressed as:

[0082]

[0083] In the formula, This is the result after the low-frequency component in the spectrum diagram F1 is replaced. Represents the data in the spectrogram F2.

[0084] In practice, the low-frequency components in the spectrum diagram F2 may also be replaced by corresponding low-frequency components in the spectrum diagram F1.

[0085] By replacing the low-frequency component, the spectrum graph F1 is changed into the spectrum graph F1′, and the spectrum graph F2 is changed into the spectrum graph F2′.

[0086] Inverse Fourier transform: Apply inverse FFT to the spectrogram F1′ to transform it back to the spatial domain and generate a new image .make sure and It is more consistent visually and reduces the impact of differences in imaging conditions on the model. The formula is:

[0087] and A new dual-phase image is formed, which is to be input into the multi-scale CNN module.

[0088] If the low-frequency components in spectrogram F2 are replaced by the corresponding low-frequency components in spectrogram F1, then an inverse FFT is applied to spectrogram F2′ to convert it back to the spatial domain, generating a new image . and A new dual-phase image is formed, which is to be input into the multi-scale CNN module.

[0089] It should be noted that the object of low-frequency component replacement of the FFT module is random. The low-frequency area of ​​the first image in the dual-phase image is randomly determined to be copied to the second image, or the low-frequency area of ​​the second image is copied to the first image. Formula 2 is used to determine which pixels to replace. By randomly applying FFT changes to the dual-phase image, the generalization ability and adaptability of the model are increased, and no replacement is required in actual detection.

[0090] This FFT module is used during training and is not replaced during actual detection.

[0091] (2) Multi-scale CNN

[0092] The multi-scale CNN uses an improved ResNet-18. ResNet-18 is a classic convolutional neural network with good feature extraction capabilities and shallow network depth, which is suitable for change detection tasks in high-resolution remote sensing images.

[0093] The multi-scale CNN is improved on the basis of the ResNet-18 structure. The improvements include: (1) Removing the final fully connected layer: focusing on multi-scale feature extraction, avoiding overfitting of specific classification tasks, and aiming to capture multi-scale features between various resblocks. (2) Retaining residual connections: Alleviating the gradient vanishing problem through skip connections, ensuring that deep networks can also be effectively trained.

[0094] like Figure 4 As shown in the left half of the figure, in this embodiment, the multi-scale CNN includes a first convolutional layer Conv1 and four residual layers Res1, Res2, Res3 and Res4 connected in sequence. During training, the first feature maps of three scales are extracted from the residual layers Res1, Res2 and Res4, which are recorded here as , , The sizes are 128×128×64, 64×64×128, and 32×32×512, where 64, 128, and 512 are the number of channels. The three feature maps are respectively input into the subsequent multi-core Inception module for processing to obtain the second feature maps of three scales. , j = 1, 2, 3. During inference prediction, these three second feature maps still enter the multi-core Inception module for subsequent processing, but only the deepest feature map is used. The obtained detection map P 3 as the final prediction result.

[0095] The first convolutional layer Conv1 uses a 7x7 convolution kernel to directly extract basic low-level features such as edges and textures from the input image. The parameters of the Conv1 layer are stride 2, padding 3, and the number of output channels is 64.

[0096] The residual block contains four residual layers Res1 to Res4, each of which is responsible for gradually extracting and refining image features, transitioning from low-level details to high-level semantic features. The residual connection alleviates the gradient vanishing phenomenon, effectively reducing the feature map size by a downsampling operation with a step size of 2, effectively halving the feature map size and expanding the receptive field. Each residual block has two paths inside, one is the ordinary convolution path, and the other is the identity mapping path. The two are added and passed to the next layer, see Figure 5 . Figure 5 The structure of a residual layer is shown. The input x jumps to the output. At the same time, the input x passes through two weight layers to generate an intermediate product F(x), which is then added to the input x and output.

[0097] (3) Multi-core Inception module (MKI)

[0098] In order to further extract multi-scale features and avoid the problems caused by using large convolution kernels or extended convolutions, FIMANet uses multiple parallel convolution kernels to extract features of different scales, which can reduce the introduction of background noise and maintain the integrity of key information.

[0099] like Figure 4 As shown in the right half of the figure, in this preferred embodiment, in the multi-core parallel convolution module, there are 5 parallel convolution branches, and the convolution kernels are 3×3, 5×5, 7×7, 9×9 and 11×11 respectively; the 5 processing results output by the parallel convolution branch are further fused through feature connection and convolution of 1×1 convolution kernel to obtain the second feature map. In the present invention, F is used to represent a feature map, the subscript i represents which path of the dual-phase image it belongs to, the superscript j represents the jth scale, and the number in the brackets represents how many times it has been processed.

[0100] The output feature map corresponding to the convolution kernel in each convolution branch It can be expressed as follows:

[0101] (5)

[0102] Assuming the input features Figure X The size is C, here Represents the convolution operation, and its convolution kernel size is , and the number of output channels is . Multiple such operations can be concatenated along the channel dimension. Then, the output feature maps of all convolution operations are concatenated along the channel dimension to produce a new feature map, namely:

[0103] (6)

[0104] Finally, the concatenated feature map The fused feature map is obtained by fusing through a 1×1×C convolutional layer. ∈ The fusion feature map The second feature map . Convolutional fusion is expressed as:

[0105] (7)

[0106] In the formula, As the weight of the 1×1 convolution kernel on channel c.

[0107] The 1×1 convolution, as a channel integration technique, merges features from different receptive field sizes, which enables the MKI module to include comprehensive contextual information while maintaining the integrity of local texture features.

[0108] (4) Cross-Scale Feature Aggregator CSFA

[0109] CSFA enhances the feature representation of the changing region through position attention, spatial attention and cross attention mechanisms.

[0110] like Figure 2 The middle part and Figure 6 As shown in the figure, CSFA consists of a position attention (PA) module, a spatial attention (SA) module, and an encoder and decoder of a Transformer module. Its function is to adaptively assign attention to multi-scale features and use Transformer to capture long-range dependencies.

[0111] The dimension of the PA module within the channel affects the second feature map Enhance the key positions and output the third feature map , the SA module performs the third feature map in the multi-channel dimension Enhance the key areas and generate feature map T; the encoder encodes the feature map T to generate context-rich tags , passed to the decoder; the decoder uses a cross-attention mechanism to generate the third feature map of the PA module as the query, and the tokens from the encoder They are used as keys and values ​​respectively to generate enhanced features, that is, to extract feature maps.

[0112] The PA module selectively enhances the key areas in the feature map, allowing the model to focus more on these salient areas in the subsequent processing stage, thereby improving the accuracy of feature extraction. Unlike the position attention module, which only focuses on the salient features in a specific channel, the SA module aims to identify key areas in the entire image globally, and generate more semantically expressive semantic tags by assigning adaptive weights to these important areas. This mechanism allows the model to model in a wider spatial context, thereby capturing global feature relationships. The Transformer Encoder (TE) is introduced to capture long-range dependencies. It can be seen that CSFA combines the processing of three dimensions, namely, position in the channel, channel-to-channel, and context, for feature enhancement. This mechanism effectively combines information from different feature sources and significantly enhances the modeling capabilities of global context and local details.

[0113] Figure 7 The structure of the PA module in the preferred embodiment of the present invention is shown. The characteristic of the structure is the design of the multi-channel point-by-point convolution layer: each output channel has its own independent convolution kernel, rather than all channels sharing the convolution kernel. Each convolution kernel is convolved only with the corresponding input channel, avoiding mixing between channels. The result size and number of channels after convolution are still the number of output channels.

[0114] As shown in Figure 7, the PA module includes a convolutional layer with a uniform number of channels, a feature adjustment module, a multi-channel point-by-point convolutional layer, a normalization module, and a weighting module, which are connected in sequence.

[0115] The convolution layer with the same number of channels is used to convert the second feature map of the input into (also simplified as X in the figure) is unified to a fixed number of channels M, generating the first intermediate feature map X0 of the M channels and outputting it to the feature adjustment module. In this embodiment, the number of channels of the unified convolution layer is The convolution operation unifies the multi-scale feature maps (with 64, 128 and 512 channels respectively) extracted in the previous stage to a fixed number of channels M = 32. This processing not only ensures the unification of the number of channels, but also retains the expressive power of multi-scale features.

[0116] The feature adjustment module includes a batch normalization layer, a linear rectification activation function, and a Dropout module, which processes the first intermediate feature map X0 to generate the second intermediate feature map X1 of M channels, further enhancing the stability and nonlinear expression ability of the features. Batch normalization, ReLU activation, and Dropout do not directly affect the generation of the weight matrix, but are used to adjust the representation of the feature map, prevent overfitting, and improve the nonlinear expression ability of the model.

[0117] The multi-channel point-by-point convolution layer is the core. This module includes M 1×1 convolution kernels, which perform convolution on the corresponding channels of the second intermediate feature map X1 without sharing the convolution kernels to generate the third intermediate feature X2. The 1×1 convolution kernel uses a 1×1 depth convolution, which is used to model each channel separately and mine more fine-grained features. The setting of this point-by-point convolution layer is that the convolution kernel size is 1 and the number of output channels is grouped as M, which means that channel-by-channel convolution is performed. In this embodiment, each output channel has its own independent convolution kernel, a total of 32, instead of all channels sharing the convolution kernel. Each convolution kernel is convolved only with the corresponding input channel to avoid mixing between channels. The size of the result and the number of channels after convolution are still the number of output channels.

[0118] The normalization module uses the Sigmoid activation function ( ) Normalize the value of each position of the third intermediate feature X2 to the range of [0, 1] and generate a position attention weight map, in which each value represents the importance of the corresponding position in the first intermediate feature map X0.

[0119] A weighting module is used to multiply the first intermediate feature map X0 and the position attention weight map element by element (⊙) to achieve position enhancement and obtain the third feature map , this feature map is a position weighted feature map, also simplified as Y in the figure.

[0120] This process of the PA module not only retains the original spatial information of the input features, but also enhances the expression ability of key areas, significantly improving the model's perception of multi-scale targets and its ability to capture local details.

[0121] Next, the feature map Y output by the PA module is passed to the SA module for further processing. Unlike the PA module, which only focuses on the salient features in a specific channel, the SA module aims to identify key areas in the entire image globally and generate more semantically expressive semantic tags by assigning adaptive weights to these important areas. This mechanism allows the model to model in a wider spatial context, thereby capturing global feature relationships.

[0122] Specifically, the process starts with steps similar to those in natural language processing, through a learnable kernel Perform point-by-point convolution on the input feature map Y. The purpose of this operation is to divide the feature map into L semantic groups to achieve grouping and representation of semantic features. This step not only retains local features, but also enhances the information separation and aggregation between different semantic groups through learned weights. Subsequently, the model applies a Softmax operation on the spatial dimension of each semantic group and calculates the corresponding spatial attention weights. These weights represent the importance of each spatial position and can effectively capture global spatial dependencies. Finally, the calculated weights are multiplied element-by-element with the input feature map Y to generate the final tag set. , this set of tags combines the global information of spatial context with the details of local salient features. The whole process can be formally expressed as:

[0123] (8)

[0124] in, (•) represents point-wise convolution, (•) denotes the Softmax operation along the channel dimension.

[0125] Finally, the Transformer Encoder (TE) is introduced to capture long-range dependencies. First, a set of trainable parameters is added element-wise to the tokens. To complete position embedding, generate context-rich tags These tokens are then passed to the Transformer Decoder (TD).

[0126] A cross-attention mechanism is used in the decoder, where the feature map generated by the PA module as the query (query, Q), and the output from the Transformer decoder Then they are used as keys ( ) and value (value, V). The corresponding formula is as follows:

[0127] (9)

[0128] in, , and denote the weight matrices of the linear layers used to project Q, K, and V into a specific representation space, respectively. This mechanism effectively combines information from different feature sources and significantly enhances the ability to model both global context and local details.

[0129] See also Figure 8 , the improved Siamese network structure of the preferred embodiment of the present invention is used for change map generation. When generating change maps, traditional methods find it difficult to take into account both local details and overall structures, resulting in inaccurate detection results. The present invention innovatively combines feature splicing and fusion: the features processed by the CSFA module are decoded by the Transformer decoder (TD), and then the decoded features are spliced ​​together and input into the prediction head to generate the final change map. Multi-scale feature integration is performed simultaneously: through the improved Siamese network structure, effective global modeling is achieved while capturing fine-grained features. The generated change map shows the detected change area, which can be directly used for analysis or decision-making assistance, providing more accurate and reliable change detection results.

[0130] (5) Output module (dual-channel merging unit)

[0131] The output module includes a connection module and a classification module. The connection module extracts the feature maps from the two paths. and Splicing; The classification module uses a convolutional layer to compare the spliced ​​feature maps to obtain the detection map P of cultivated land changes.

[0132] During training, the multi-scale feature maps generated by the multi-scale CNN are processed by the MKI, CSFA modules and output modules to form multi-scale detection maps P j In actual detection reasoning, only the detection map corresponding to the feature map output by the last residual layer of the multi-scale CNN is used as the detection result.

[0133] (6) Training unit

[0134] The multi-scale training module in the training unit uses the detection map P of all scales j When calculating the loss jointly with the true value, the binary cross entropy (BCE) loss is used as the objective function. BCE loss is one of the most widely used loss functions for binary classification and image segmentation because it can assign different weights to each pixel and evaluate the similarity of dual-phase images based on spatial information. The definition of BCE loss is as follows:

[0135] (10)

[0136] Where N is the total number of pixels in the image. It represents the true label of the i-th pixel in the real image, and its value is 0 or 1. To detect the probability that the i-th pixel in the image P is predicted to be category 1, the predicted value is usually calculated by the sigmoid activation function.

[0137] , and denote the output detection images of different scales respectively, let G denote the ground truth, and the total loss function of FIMANet uses the cross entropy ( ) The loss formula is as follows:

[0138] (11)

[0139] The effect of the present invention is verified below.

[0140] 1. Dataset selection

[0141] To verify the effectiveness of the present invention, three challenging change detection datasets were selected: CLCD, WHU-CD, and LEVIR-CD. The CLCD dataset focuses on farmland changes and contains 600 pairs of dual-phase images of 512x512 pixels with a resolution of 0.5-2 meters. The dataset is divided into training, validation, and test sets, which are suitable for evaluating change detection algorithms in the agricultural field. WHU-CD and LEVIR-CD focus on building change detection, respectively containing 1 pair of 32,507×15,354 pixels and 637 pairs of 1024×1024 pixels dual-phase images, with spatial resolutions of 0.2 meters and 0.5 meters, respectively. The data of both are cropped to generate small blocks of 256×256 pixels and randomly divided into training, validation, and test sets, which are suitable for evaluating algorithm performance.

[0142] 2. Experimental environment construction

[0143] The experimental environment is based on PyTorch 2.2.2 and CUDA 11.8, and the deep learning model is trained on an NVIDIA RTX 4090 GPU. The Adam optimizer is used, the initial learning rate is 0.0001, the beta parameter is (0.9, 0.999), and the cosine annealing scheduler is used to dynamically adjust the learning rate, with the minimum value set to 1e-5. The training lasts for 100 epochs, and validation is performed after each epoch. The model with the highest intersection over union (IoU) on the validation set is selected as the final model.

[0144] 3. Dataset training

[0145] The datasets (CLCD, WHU-CD, LEVIR-CD) were cropped, normalized, and augmented in the preprocessing stage to improve the generalization ability of the model. The preprocessed datasets were divided into training, validation, and test sets. During the training process, batch gradient descent was used to optimize the model, and the learning rate scheduler and early stopping method were used to prevent overfitting. At the same time, cross-validation was used to evaluate the performance stability of the model on different datasets.

[0146] 4. Experimental method selection

[0147] In order to verify the effectiveness of the present invention, a variety of advanced change detection methods were selected for comparison, including: FC-EF (based on U-Net architecture, extracting joint features through image-level fusion); FC-Siam-conc (based on Siamese network, processing dual-phase images and extracting features); FC-Siam-diff (calculating pixel-by-pixel differences to capture changes); DTCDSCN (combining channel and spatial attention mechanisms to enhance features); SNUNet (combining Siamese networks with Nested-Unet for multi-scale feature extraction); BIT (combining CNN and transformer to capture contextual information using self-attention); MSCANet (integrating multi-scale context aggregator and multi-branch prediction head); SARASNet (introducing relationship perception and scale sensitivity to solve scene change problems); HATNet (combining self-attention and cross-attention mechanisms for building change detection).

[0148] 5. Results comparison

[0149] In the result comparison stage, a variety of evaluation indicators were used to quantify the performance of different methods, including precision, recall, F1 score, intersection over union (IoU) and overall accuracy (OA). The experimental results show that the proposed scheme has achieved significantly better performance than the baseline method on all three datasets.

[0150] (1) Quantitative experimental comparison

[0151] Table 1 summarizes the quantitative results of each model on the CLCD, WHU-CD and LEVIR-CD test sets. FIMANet outperforms all the compared models in all five indicators.

[0152] Table 1

[0153]

[0154] In the table above, the best scores are shown in bold and all results are expressed as percentages (%).

[0155] (3) Comparison of qualitative results

[0156] Fig. 9 and Fig.10The prediction result map of the model on different datasets is shown, and different color codes are used to present the detection results: true positive examples (TP) are represented by white, true negative examples (TN) are represented by black, false positive examples (FP) are represented by purple, and false negative examples (FN) are represented by green. This color mapping method intuitively highlights the accuracy and error types of model predictions. The fewer green and purple areas in the figure indicate that the algorithm performs better in recall and precision, and can more effectively identify real change areas and reduce false detections. The experimental results show that the proposed FIMANet achieves SOTA (state-of-the-art) performance on all three datasets.

[0157] (4) Frequency domain sampling point parameter setting and visualization in FFT

[0158] To evaluate the effect of the frequency domain sampling size ∆ on the formula, the formula was verified by gradually increasing the sampling point size by 11 (with a step size of 5 sampling points). Table 2 shows that when the frequency domain sampling point increases from 0 to 10, the F1 score and IoU on the CLCD, WHU-CD and LEVIR-CD datasets gradually increase, indicating that FFTFDFI effectively enhances the information coupling of the dual-phase images and reduces the interference of pseudo changes. However, as the interaction area is further expanded, the F1 score and IoU decrease because the excessive replacement leads to the loss of source image content and the generation of artifacts. Therefore, it is most appropriate to set ∆ to 10.

[0159] To further analyze the impact, Fig.11 (a) shows the image after FFT processing. In order to further analyze the impact, Fig.11 (a) shows the image after FFT processing. Visually with Closer. Fig.11 (b) in the figure shows the red band frequency domain diagram of the nine images, and the relative mean energy spectral density (RMESD) is marked in the upper left corner of the figure. After logarithmic transformation and normalization, The RMESD values ​​(10548, 9349, 10413) are closer (9956, 8928, 10108), and In the frequency domain, the information of the image is decomposed into sine and cosine waves of different frequencies, so that features or patterns of different scales in the image can be observed, such as textures, periodic structures, etc. Fig.11 The frequency domain image in (b) shows the distribution of the original image at different frequency components. The central symmetric pattern represents zero frequency, and the points far from the center correspond to high-frequency information. Frequency domain analysis helps to reveal the spatial variation characteristics of the image, which is crucial for image processing tasks such as filtering and feature extraction.

[0160] Table 2

[0161]

[0162] In the table above, the best scores are shown in bold and all results are expressed as percentages (%).

[0163] (5) CSFA module visualization

[0164] In order to understand the CSFA module more intuitively, an example is selected from the CLCD dataset and multiple representative dual-phase feature maps of each stage are visualized. Fig.12 As shown, the multi-scale feature map extracted by the MKI module and . Fig.12 (b) shows clear boundary contours and highly compact internal structures, but the model shows relatively low global attention. Subsequently, the PA module is and Assign pixel-by-pixel attention weights and generate feature maps and ( Fig.12 In these images, we can observe the obvious contour features of key areas of interest, such as individual farmland plots and backfill areas. Then, the feature map obtained by the SA module and ( Fig.12 (d) further enhances the expression of regional variation features, allowing the model to more effectively aggregate meaningful spatial features. Finally, the Transformer module obtains and ( Fig.12 (e)) demonstrates a sophisticated understanding of high-level semantic relationships. For example, in In , all farmland areas are closely clustered together, while in In the figure, the boundary between farmland and backfill soil is clearly distinguished. The final detection result figure also verifies the effectiveness of the CSFA module in accurately allocating attention to the bi-temporal feature map.

[0165] The above specific embodiments only describe the design principle of the present invention. The shapes and names of the components in the description may be different and are not limited. Therefore, those skilled in the art in the field of the present invention may modify or replace the technical solutions recorded in the above embodiments; and these modifications and replacements do not deviate from the creative purpose and technical solutions of the present invention and should all fall within the protection scope of the present invention.

Claims

1. A multi-scale farmland change detection system based on frequency domain conversion, the system comprising a feature extraction model, a dual-path merging unit and a training unit; characterized in that: The feature extraction model processes each channel of the input dual-phase image to obtain two extracted feature maps and ; The feature extraction model includes a feature extraction module, a multi-core parallel convolution module, and a feature enhancement module; A feature extraction module is used to extract features from two images in the dual-phase image to obtain a multi-scale first feature map; The multi-core parallel convolution module contains multiple parallel convolution branches. The parallel convolution branches use m convolution kernels of different sizes, and the convolution kernel is less than or equal to 13×13; the value range of m is 3-7; the first feature map enters each parallel convolution branch for processing, and then is spliced ​​and fused to obtain the second feature map ; Feature enhancement module, used to enhance the second feature map Perform feature enhancement and generate extracted feature maps ; The two-way merging unit extracts feature maps from two paths and Merge and classify to obtain the detection map P of cultivated land changes; The training unit includes an FFT module and a multi-scale training module; The FFT module is used to convert the bi-phase image from the spatial domain to the frequency domain when training the feature extraction model, unify the low-frequency areas in the bi-phase image by replacement operation, and then convert it back to the spatial domain to generate a bi-phase image after the low-frequency component is replaced as the bi-phase image used for training; When the multi-scale training module trains the feature extraction model, the multi-scale first feature map from the feature extraction module is processed by the multi-core parallel convolution module, the feature enhancement module and the dual-channel merging unit to generate a multi-scale detection map P j , j represents the jth scale; using the detection map P of all scales j The loss is calculated jointly to optimize the feature extraction model. During inference detection, only the detection map P corresponding to the first feature map of the deepest layer is used as the inference result.

2. The multi-scale farmland change detection system based on frequency domain conversion according to claim 1, characterized in that: When the FFT module replaces the low frequency area of ​​the dual-phase image, for a series of dual-phase images, it randomly determines to copy the low frequency area of ​​the first image in the dual-phase image to the second image, or to copy the low frequency area of ​​the second image to the first image.

3. The multi-scale farmland change detection system based on frequency domain conversion according to claim 1, characterized in that: The feature extraction module uses the ResNet-18 backbone network with the fully connected layer removed.

4. The multi-scale farmland change detection system based on frequency domain conversion according to claim 3 is characterized in that: The feature extraction module includes a first convolution layer Conv1 and four residual layers Res1, Res2, Res3 and Res4 connected in sequence; during training, first feature maps of three scales are extracted from the residual layers Res1, Res2 and Res4, and are respectively input into the multi-core parallel convolution module for processing to obtain second feature maps of three scales; during inference prediction, only the deepest first feature map output by the residual layer Res4 is used to input into the multi-core parallel convolution module for processing.

5. The multi-scale farmland change detection system based on frequency domain conversion according to claim 4 is characterized in that: In the feature extraction module, the first convolution layer Conv1 uses a 7×7 convolution kernel, and the residual layers Res2 and Res3 are downsampled with a step size of 2.

6. The multi-scale farmland change detection system based on frequency domain conversion according to claim 1, characterized in that: In the multi-core parallel convolution module, there are 5 parallel convolution branches, and the convolution kernels are 3×3, 5×5, 7×7, 9×9 and 11×11 respectively; the 5 processing results output by the parallel convolution branches are further fused through feature connection and convolution of the 1×1 convolution kernel to obtain a second feature map.

7. The multi-scale farmland change detection system based on frequency domain conversion according to claim 1, characterized in that: The feature enhancement module enhances the feature representation of the changing area through position attention, spatial attention and cross attention mechanisms.

8. The multi-scale farmland change detection system based on frequency domain conversion according to claim 7, characterized in that: The feature enhancement module adopts a cross-scale feature aggregator; The cross-scale feature aggregator consists of a position attention PA module, a spatial attention SA module, and the encoder and decoder of the Transformer module; The dimension of the PA module within the channel affects the second feature map Enhance the key positions and output the third feature map ; The SA module performs the third feature map in the multi-channel dimension Enhance the key areas and generate feature map T; the encoder encodes the feature map T to generate a tag with context , passed to the decoder; The decoder uses a cross-attention mechanism to generate the third feature map of the PA module. As query, the tokens from the encoder As key and value respectively, generate enhanced features, that is, extract feature maps .

9. The multi-scale farmland change detection system based on frequency domain conversion according to claim 8, characterized in that: The PA module includes a convolution layer with a uniform number of channels, a feature adjustment module, a multi-channel point-by-point convolution layer, a normalization module and a weighting module connected in sequence; The number of channels is unified in the convolution layer, which is used to convert the second feature map of the input into Unify to a fixed number of channels M, generate the first intermediate feature map X0 of M channels and output it to the feature adjustment module; The feature adjustment module includes a batch normalization layer, a linear rectification activation function and a Dropout module, processes the first intermediate feature map X0, and generates a second intermediate feature map X1 of M channels; The multi-channel point-by-point convolution layer includes M 1×1 convolution kernels, which perform convolution on corresponding channels of the second intermediate feature map X1 without sharing the convolution kernels, and generate the third intermediate feature X2; The normalization module uses a Sigmoid activation function to normalize the value of each position of the third intermediate feature X2 to the range of [0,1] to generate a position attention weight map; each value in the position attention weight map represents the importance of the corresponding position in the first intermediate feature map X0; The weighting module is used to perform element-by-element multiplication of the first intermediate feature map X0 and the position attention weight map to achieve position enhancement and obtain the third feature map .

10. The multi-scale farmland change detection system based on frequency domain conversion according to claim 1, characterized in that: The dual-path merging unit includes a connection module and a classification module; the connection module extracts feature maps from the two paths and Splicing; The classification module uses a convolutional layer to compare the concatenated feature maps to obtain a detection map of cultivated land changes.

Images