Remote sensing image change detection method and system based on Fourier frequency domain feature enhancement

Abstract

The invention discloses a remote sensing image change detection method and system based on Fourier frequency domain feature enhancement. The method comprises the following steps: acquiring double-time-phase images: a first image and a second image; performing shallow convolution flow on the first image and the second image through a shared weight to obtain a first initial feature map and a second initial feature map; respectively inputting the first initial feature map and the second initial feature map into a corresponding multi-cascade encoder to obtain a first deep feature and a second deep feature; inputting the first deep feature and the second deep feature into a self-adaptive dual-time-phase interaction module, and performing cross-time-phase interaction and guide fusion to generate a multi-scale difference feature; and inputting the multi-scale difference characteristics into a decoder to obtain a change probability graph. According to the method, the contradiction between global modeling efficiency and local detail reservation of the existing architecture can be overcome, explicit and efficient fusion of the local spatial features and the global frequency domain features is realized, and the change detection precision and efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of high-resolution remote sensing image change detection technology, and in particular to a remote sensing image change detection method and system based on Fourier frequency domain feature enhancement. Background Technology

[0002] Remote sensing image change detection aims to identify areas of change on the Earth's surface from observational data across different time periods, playing a crucial role in applications such as urban sprawl monitoring, ecological environment assessment, and post-disaster damage detection. However, with the widespread acquisition of high-resolution optical imagery, change detection methods are required to simultaneously capture fine local structural details and complex global change patterns. Nevertheless, effectively balancing local and global feature modeling remains a core challenge within existing deep learning frameworks.

[0003] In the field of remote sensing change detection, convolutional neural network (CNN)-based models have long been the mainstream approach. Typical architectures include FC-EF, FC-Siam-Conc, and FC-Siam-Diff. FC-EF stitches together bi-temporal images channel-by-channel as the overall input. FC-Siam-Conc and FC-Siam-Diff, on the other hand, use the Siamese architecture to process bi-temporal images separately, then perform fusion or difference calculations at the feature layer. Subsequent works have extended this approach, employing strategies such as densely connected CNNs, multi-level supervision, attention mechanisms, differential enhancement modules, and multi-scale weighted fusion to improve boundary precision, robustness, and class discriminative power in change detection. Transformer-based or self-attention mechanisms establish long-distance spatial relationships through global attention and enhance the interaction between features from different temporal phases. For example, some models insert Transformer layers into the Siamese architecture to enhance global context modeling. Other models construct pure Transformer networks for change detection, combined with multi-level attention fusion mechanisms. Hybrid architectures (such as CNN+Transformer) use CNNs to extract local features in shallow or mid-level layers and Transformers to capture global semantics in high-level layers. Furthermore, to address the specific needs of change detection, some methods incorporate spatiotemporal relationship modeling and cross-scale interaction mechanisms into their attention design to improve the ability to discriminate changed regions. Structured State Space Models (SSMs) and their efficient implementation, Mamba, have recently gained attention in the fields of vision and sequence modeling. Their linear complexity in input length and powerful global modeling capabilities make them one of the alternatives to Transformers. In vision tasks, versions such as Vision Mamba and VMamba (four-way scanning) have been proposed to adapt to image inputs with two-dimensional spatial structures. Against this backdrop, some methods apply the Mamba architecture to remote sensing change detection. For example, ChangeMamba uses VMamba as the encoder backbone, achieves global image context modeling through cross-scan modules, and designs cross-spatiotemporal state space modules to capture the spatiotemporal correlation between bi-temporal features. Some extended versions (such as RS-Mamba and DC-Mamba) enhance boundary cues and local feature representations by improving scanning strategies, introducing edge enhancement modules, or dual-stream state space structures. Building on this, CDMamba introduces local convolutional branches into the pure Mamba architecture, consciously fusing global and local features to compensate for the shortcomings of traditional Mamba methods in detecting boundary details and subtle changes.The advantage of Mamba-based methods lies in their ability to model global semantics with linear complexity under high-resolution input, thereby alleviating the overhead bottleneck of Transformers. Their state evolution mechanism allows for information transfer across locations and time phases, making them suitable for capturing long-range dependencies.

[0004] CNN-based methods excel at capturing local spatial features, but their limited receptive field restricts their ability to model long-range dependencies, resulting in insufficient representation of global structural information. Transformer-based models achieve global dependency modeling through self-attention mechanisms, but their computational complexity increases quadratically with image resolution and is susceptible to spatial redundancy in high-resolution remote sensing imagery, leading to high computational costs. State-space models (such as Mamba), which have emerged in recent years, demonstrate superior efficiency in sequence modeling, but their inherent directional scanning mechanism leads to biases in their global modeling capabilities in two-dimensional space, making it difficult to achieve non-directional global feature representation. Furthermore, from the perspective of the specific task of detecting changes in dual-temporal remote sensing images, different architectures exhibit distinctly different feature response patterns. CNN-based architectures rely on local convolutional kernels for edge detection and texture analysis, generating spatially localized high-response regions, primarily capturing local variation patterns in images. Transformer-based architectures utilize global context modeling mechanisms, employing multi-head attention to model long-range spatial dependencies, resulting in smooth and globally consistent feature distributions. Mamba-based architectures simulate the sequential scanning characteristics of state-space models, combining bidirectional state propagation and selective gating mechanisms to exhibit sequential and directional responses in feature maps. However, none of these three approaches can simultaneously achieve locality, globality, efficiency, and non-directionality in feature modeling. Most importantly, the more complex dual-temporal high-resolution remote sensing images possess rich local spatial details and non-directional global structural information; efficiently and accurately capturing these features requires models with a theoretically higher balance. Summary of the Invention

[0005] This invention provides a remote sensing image change detection method and system based on Fourier frequency domain feature enhancement, which can overcome the contradiction between global modeling efficiency and local detail preservation in the existing architecture, realize explicit and efficient fusion of local spatial features and global frequency domain features, and improve the accuracy and efficiency of change detection.

[0006] This invention provides a method for detecting changes in remote sensing images based on Fourier frequency domain feature enhancement, comprising: Acquire dual-temporal images: the first image and the second image; The first image and the second image are passed through a shallow convolutional stream with shared weights to obtain a first initial feature map and a second initial feature map; The first initial feature map and the second initial feature map are respectively input into the corresponding multi-cascade encoder to obtain the first deep feature and the second deep feature; The first deep feature and the second deep feature are input into the adaptive dual-temporal interaction module to perform cross-temporal interaction and guided fusion, generating multi-scale difference features; The multi-scale difference features are input into the decoder to obtain the change probability map.

[0007] Furthermore, according to the above-mentioned remote sensing image change detection method based on Fourier frequency domain feature enhancement, the two multi-cascade encoders have the same structure, both including multiple frequency domain feature enhancement hybrid extraction modules connected in sequence, and each frequency domain feature enhancement hybrid extraction module includes local feature paths and global frequency domain feature paths. The processing procedure of the multi-cascade encoder includes: After performing layer normalization on the initial feature map, a normalized initial feature map is obtained. The normalized initial feature map is then input into the local feature path and the global frequency domain feature path to obtain local features and global frequency domain features, respectively. The local features and the global frequency domain features are fused and then linearly mapped to obtain the mapped features. The mapped features and the initial feature map are then merged through scale residual connection to obtain the deep features.

[0008] Furthermore, according to the above-described remote sensing image change detection method based on Fourier frequency domain feature enhancement, the processing of the local feature path includes: The normalized initial feature map is passed through a two-dimensional convolutional layer and then through an activation function to obtain local features.

[0009] Furthermore, according to the above-mentioned remote sensing image change detection method based on Fourier frequency domain feature enhancement, the processing procedure of the global frequency domain feature path includes: Flatten the normalized initial feature map to obtain the feature sequence; Perform a discrete Fourier transform on the feature sequence to obtain a complex frequency spectrum, and decompose the complex frequency spectrum into a real part and an imaginary part; The real and imaginary parts are recombined, and the spatiotemporal sequence is reconstructed back using inverse Fourier transform; The spatiotemporal sequence is linearly mapped to obtain global frequency domain features.

[0010] Furthermore, according to the aforementioned remote sensing image change detection method based on Fourier frequency domain feature enhancement, the adaptive dual-temporal interaction module includes parallel sub-modules: a long-range context interaction unit and a fine spatial alignment module; the processing procedure of the adaptive dual-temporal interaction module includes: Using the second deep feature as the guiding feature and the first deep feature as the guided feature, the first deep feature and the second deep feature are input into the long-range context interaction unit of the submodule to obtain the first global guiding feature. The first deep feature and the second deep feature are input into the fine spatial alignment module to obtain the first local spatial feature. Using the first deep feature as the guiding feature and the second deep feature as the guided feature, the first deep feature and the second deep feature are input into the long-range context interaction unit of the submodule to obtain the second global guiding feature. The first deep feature and the second deep feature are input into the fine spatial alignment module to obtain the second local spatial feature. Based on a dynamic gating mechanism, the first global guidance feature and the first local spatial feature are fused to obtain a first fused feature, and the second global guidance feature and the second local spatial feature are fused to obtain a second fused feature; Multi-scale difference features are obtained based on the first fusion feature and the second fusion feature.

[0011] Furthermore, according to the above-mentioned remote sensing image change detection method based on Fourier frequency domain feature enhancement, the second deep feature is used as the guiding feature, and the first deep feature is used as the guided feature. The first deep feature and the second deep feature are input into the long-range context interaction unit of the sub-module to obtain the first global guiding feature. The first deep feature and the second deep feature are input into the fine spatial alignment module to obtain the first local spatial feature, including: The first deep feature is obtained by applying a linear transformation and an activation function to the first feature. The first deep feature is input into the selective state space model after linear transformation and convolutional layer to obtain the first context information. The first context information is then normalized by layer to obtain the normalized first context information. The second feature is obtained by fusing the first feature and the normalized first context information; The second deep feature is processed through a linear transformation and a convolutional layer and then input into a selective state space model to obtain the corresponding second context information. The second context information is then processed through an activation function to obtain the transformed second context information. The second feature, the normalized first context information, and the transformed second context information are fused to obtain the first global guiding feature; The third feature is obtained by applying a linear transformation and activation function to the first deep feature; The first deep feature is input into the selective state space model after linear transformation and convolutional layer to obtain the third context information. The third context information is then normalized by layer to obtain the normalized third context information. The fourth feature is obtained by fusing the third feature and the normalized third context information; The second deep feature is passed sequentially through a convolutional layer, an activation function, another convolutional layer, and another activation function to obtain the fifth feature; The fourth feature, the normalized third contextual information, and the fifth feature are fused to obtain the first local spatial feature.

[0012] Furthermore, according to the above-mentioned remote sensing image change detection method based on Fourier frequency domain feature enhancement, wherein, based on a dynamic gating mechanism, the first global guiding feature and the first local spatial feature are fused to obtain a first fused feature, and the second global guiding feature and the second local spatial feature are fused to obtain a second fused feature, including: The first global guiding feature and the first local spatial feature are compressed along the spatial dimension and then concatenated. Then, through a linear layer and an activation function, two sets of first gating scores are obtained. The first global guidance feature and the first local spatial feature are weighted and fused using the two sets of first gating scores to obtain the first fused feature; The second global guiding feature and the second local spatial feature are compressed along the spatial dimension and then concatenated. Then, through a linear layer and an activation function, two sets of second gating scores are obtained. The second global guiding feature and the second local spatial feature are weighted and fused using the set of second gating scores to obtain the second fused feature.

[0013] Furthermore, according to the aforementioned remote sensing image change detection method based on Fourier frequency domain feature enhancement, the formula for calculating the two sets of first gate scores is as follows:

[0014] in, The first gate score, Indicates global guidance features. Representing local spatial features, For activation function, For linear layers, For splicing operations, This indicates compression along the spatial dimension. The above equation yields two sets of normalized first-gated scores. and .

[0015] Furthermore, according to the above-mentioned remote sensing image change detection method based on Fourier frequency domain feature enhancement, the first global guiding feature and the first local spatial feature are weighted and fused using the two sets of first gating scores to obtain the first fused feature, which is expressed by the following formula:

[0016] in, For element-wise multiplication, This is the first fusion feature.

[0017] This invention also provides a remote sensing image change detection system based on Fourier frequency domain feature enhancement, comprising: The acquisition module is used to acquire dual-temporal images: a first image and a second image; The feature extraction module is used to pass the first image and the second image through a shallow convolutional stream with shared weights to obtain a first initial feature map and a second initial feature map; A multi-cascade encoder is used to input the first initial feature map and the second initial feature map into the corresponding multi-cascade encoder to obtain the first deep feature and the second deep feature. An adaptive dual-temporal interaction module is used to perform cross-temporal interaction and guided fusion of the first deep feature and the second deep feature to generate multi-scale difference features; A decoder is used to decode the multi-scale difference features to obtain a change probability map.

[0018] This invention provides a remote sensing image change detection method and system based on Fourier frequency domain feature enhancement. By designing a novel dual-stream FourierMamba change detection framework, this invention overcomes the contradiction between global modeling efficiency and local detail preservation in existing architectures, achieving explicit and efficient fusion of local spatial features and global frequency domain features, thus improving change detection accuracy and efficiency. Compared with existing technologies, this invention has the following advantages: (1) A new change detection framework based on Fourier transform is proposed. This framework retains the traditional end-to-end structure, and not only has the feature information mining capabilities of local perception and global modeling, but also takes into account higher work efficiency. It is a new paradigm for change detection. (2) A frequency domain feature enhancement hybrid extraction module, FHEB, is proposed, which combines the traditional local convolution extraction path with a global frequency domain analysis path based on Fast Fourier Transform (FFT). This design allows the model to efficiently capture long-range dependencies in a non-directional and holistic manner through the inherent characteristics of the frequency domain, while avoiding the dependence on scanning patterns in the Mamba architecture; (3) A mathematically rigorous frequency domain feature path was constructed. Within FHEB, the complex spectrum after FFT is innovatively decomposed into a real part (amplitude information) and an imaginary part (phase information), and these are learned and transformed independently. This method ensures that the network can perform detailed and targeted modeling of the two key frequency domain elements: global structure and spatial positioning. (4) An adaptive dual-temporal interaction module (ABTIM) was designed to dynamically fuse global and local enhancement features from two temporal images to extract the most discriminative difference features, thereby improving the accuracy of change detection. Attached Figure Description

[0019] The technical solution and other beneficial effects of the present invention will become apparent from the following detailed description of specific embodiments of the invention, in conjunction with the accompanying drawings.

[0020] Figure 1 The flowchart illustrates a remote sensing image change detection method based on Fourier frequency domain feature enhancement, as provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the frequency domain feature enhancement hybrid extraction module provided in an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the adaptive dual-time phase interaction module provided in an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the structure of a remote sensing image change detection system based on Fourier frequency domain feature enhancement provided in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figure 1 , Figure 1 The flowchart below illustrates a remote sensing image change detection method based on Fourier frequency domain feature enhancement, as provided in this embodiment of the invention. This method includes the following steps: S1, Obtain dual-temporal images: First image Second image , .

[0026] S2, passing the first image and the second image through a shallow convolutional stream with shared weights to obtain the first initial feature map and the second initial feature map.

[0027] First image Second image Shallow convolutional streams with shared weights Shallow convolutional flows with shared weights The image is converted from the RGB color space to the feature space to obtain the first initial feature map. Second initial feature map , .

[0028] S3, input the first initial feature map and the second initial feature map into the corresponding multi-cascade encoder to obtain the first deep feature and the second deep feature.

[0029] Specifically, the two multi-cascade encoders have the same structure, each including multiple frequency domain feature enhancement hybrid extraction modules (FHEB) connected in sequence. Each frequency domain feature enhancement hybrid extraction module includes a local feature path (LFP) and a global frequency domain feature path (GFFP). Figure 2 This is a schematic diagram of the frequency domain feature enhancement hybrid extraction module provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the processing steps of a multi-cascade encoder include: S31. After performing layer normalization on the initial feature map, a normalized initial feature map is obtained. The normalized initial feature map is then input into the local feature path and the global frequency domain feature path to obtain local features and global frequency domain features.

[0030] Specifically, for the initial feature map (the first initial feature map) Or the second initial feature map After performing layer normalization (LayerNorm), a normalized initial feature map is obtained. The normalized initial feature map is then input into the local feature path to obtain local features, and the normalized initial feature map is then input into the global frequency domain feature path to obtain global frequency domain features.

[0031] S32, after fusing local features and global frequency domain features, obtains mapped features through linear mapping. The mapped features and the initial feature map are then merged through scale residual connection to obtain deep features.

[0032] Steps S31-S32 are expressed by the following formula:

[0033]

[0034] in, For mapping features, For linear mapping, For layer normalization, The first initial feature map as input Or the second initial feature map , These are learnable scaling coefficients. This refers to deep features (either the first or second deep feature).

[0035] It should be noted that the input initial feature map of the cascade encoder is either the first initial feature or the second initial feature. When the input is the first initial feature, the output is the first deep feature; when the input is the second initial feature, the output is the second deep feature. The processing flow of the two multi-cascade encoders is the same, so it will not be described separately here.

[0036] In one embodiment, the processing of Local Feature Path (LFP) includes: S311, the normalized initial feature map is passed through a 3×3 two-dimensional convolutional layer (Conv2d) and then through the SiLU activation function to obtain local features. .

[0037] Local features before being fused with the output of GFFP It will be flattened into a one-dimensional sequence ,in .

[0038] In one embodiment, the processing of the global frequency domain feature path includes: S312, flatten the normalized initial feature map to obtain the feature sequence.

[0039] S313, perform a discrete Fourier transform on the feature sequence to obtain the complex frequency spectrum, and decompose the complex frequency spectrum into real and imaginary parts.

[0040] Specifically, the normalized initial feature map is converted into a feature sequence. Then, an FFT operation is performed to obtain the complex frequency spectrum. Then, perform independent linear transformations. Decompose into real part and the virtual part To precisely control the enhancement of global structure and spatial positioning information, independent linear transformations are applied respectively:

[0041] S314 reassembles the real and imaginary parts and reconstructs the spatiotemporal sequence using inverse Fourier transform.

[0042] Perform iFFT reconstruction to transform the data. and Reorganized into a plural And apply inverse FFT Reconstruct it back into a spatiotemporal sequence .

[0043] S315 performs a linear mapping on the spatiotemporal domain sequence to obtain global frequency domain features.

[0044] Final spatiotemporal sequence After linear mapping, the global frequency domain features are obtained. .

[0045] S4 inputs the first and second deep features into the adaptive dual-temporal interaction module to perform cross-temporal interaction and guided fusion, generating multi-scale difference features.

[0046] Figure 3 This is a schematic diagram of the structure of the adaptive dual-temporal interaction module provided in an embodiment of the present invention, as shown below. Figure 3 As shown, the adaptive dual-temporal interaction module includes parallel sub-modules: the Long-Range Contextual Interaction Unit (LRCIU) and the Fine-Grained Spatial Alignment Module (FSAM). They enhance the feature representation of another image (the first deep feature) by using features of one image as a guide (e.g., guiding the first deep feature through the second deep feature).

[0047] The LRCII unit focuses on leveraging contextual information enhanced by the global frequency domain for interaction. It employs a sequence operation-based mechanism (Conv1D and SSM-like structure) to extract global guiding features from the second deep feature layer. These guiding features are then passed through a Sigmoid gating mechanism. FSAM operates on the target features extracted from the first deep feature layer to achieve feature calibration based on global context. It focuses on pixel-level local feature alignment. It utilizes 2D convolution (Conv2d) to extract precise local spatial features from the second deep feature layer and applies them to the feature map of the first deep feature layer through a gating mechanism. This mechanism ensures that subtle spatial variation details are preserved and emphasized during feature difference extraction.

[0048] The processing steps of the adaptive dual-time phase interaction module include: S41, using the second deep feature as the guiding feature and the first deep feature as the guided feature, input the first deep feature and the second deep feature into the long-range context interaction unit of the submodule to obtain the first global guiding feature, and input the first deep feature and the second deep feature into the fine spatial alignment module to obtain the first local spatial feature.

[0049] Please continue reading. Figure 3 Step S41 specifically includes the following steps: S411, the first deep feature is obtained by applying the linear transformation Linear and the activation function SiLU to the first feature; S412, the first deep feature is input into the selective state space model SSM after passing through the linear transformation Linear and the convolutional layer Conv1d to obtain the first context information. The first context information is then normalized by passing through the layer normalization LayerNorm to obtain the normalized first context information. S413, the first feature and the normalized first context information are fused to obtain the second feature; S414, the second deep feature is passed through Linear and Conv1d convolutional layers and then input into the selective state space model SSM to obtain the corresponding second context information. The second context information is then passed through the activation function Sigmoid to obtain the transformed second context information. S415, the second feature, the normalized first context information and the transformed second context information are fused to obtain the first global guiding feature; S416, the first deep feature is processed by Linear and the activation function SiLU to obtain the third feature; S417, the first deep features are input into the selective state space model SSM after passing through Linear and Conv1d convolutional layers to obtain the third context information. The third context information is then normalized by LayerNorm to obtain the normalized third context information. S418, the third feature and the normalized third contextual information are fused to obtain the fourth feature; S419, the second deep feature is passed through a convolutional layer Conv2d, an activation function SiLU, another convolutional layer Conv2d, and an activation function Sigmoid in sequence to obtain the fifth feature; S41A, the fourth feature, the normalized third contextual information, and the fifth feature are fused to obtain the first local spatial feature.

[0050] S42, using the first deep feature as the guiding feature and the second deep feature as the guided feature, input the first deep feature and the second deep feature into the long-range context interaction unit of the submodule to obtain the second global guiding feature, and input the first deep feature and the second deep feature into the fine spatial alignment module to obtain the second local spatial feature.

[0051] Step S42 is similar to S41, except that the positions of the first deep feature and the second deep feature are swapped, which will not be described in detail here.

[0052] S43, based on the dynamic gating mechanism, the first global guidance feature and the first local spatial feature are fused to obtain the first fused feature, and the second global guidance feature and the second local spatial feature are fused to obtain the second fused feature.

[0053] Step S43 includes the following steps: S431: The first global guiding feature and the first local spatial feature are compressed (averaged) along the spatial dimension and then concatenated along the channel dimension. Then, through a linear layer and the Softmax activation function, two sets of first gating scores are obtained.

[0054]

[0055] in, The first gate score, Indicates global guidance features. Representing local spatial features, For activation function, For linear layers, For splicing operations, This indicates compression along the spatial dimension. The above equation yields two sets of normalized first-gated scores. and .

[0056] S432, the first fused feature is obtained by weighted fusing the first global guiding feature and the first local spatial feature through two sets of first gating scores. This is expressed by the following formula:

[0057] in, For element-wise multiplication, This is the first fusion feature.

[0058] S433 compresses the second global guiding feature and the second local spatial feature along the spatial dimension and then concatenates them. Then, through a linear layer and an activation function, two sets of second gating scores are obtained.

[0059] Specifically, two sets of second gating scores can be obtained using the formula provided in step S431. and .

[0060] S434, the second global guiding feature and the second local spatial feature are weighted and fused by a set of second gating scores to obtain the second fused feature.

[0061] Similarly, the second fusion feature can be obtained using the formula provided in step S432. .

[0062] S44, multi-scale difference features are obtained based on the first fusion feature and the second fusion feature.

[0063] Subtracting the first and second fusion features and taking the absolute value yields the multi-scale difference features. It can be expressed by the following formula:

[0064] S5 inputs the multi-scale difference features into the decoder to obtain the change probability map.

[0065] Specifically, the decoder performs step-by-step upsampling and fusion, and finally outputs a pixel-level probability map of change through a linear projection layer.

[0066] Based on the methods described in the above embodiments, this embodiment will be further described from the perspective of a remote sensing image change detection system based on Fourier frequency domain feature enhancement.

[0067] Please see Figure 4 , Figure 4 The invention specifically describes a remote sensing image change detection system based on Fourier frequency domain feature enhancement, which may include: The acquisition module is used to acquire dual-temporal images: a first image and a second image; The feature extraction module is used to pass the first image and the second image through a shallow convolutional stream with shared weights to obtain a first initial feature map and a second initial feature map; A multi-cascade encoder is used to input the first initial feature map and the second initial feature map into the corresponding multi-cascade encoder to obtain the first deep feature and the second deep feature. An adaptive dual-temporal interaction module is used to perform cross-temporal interaction and guided fusion of the first deep feature and the second deep feature to generate multi-scale difference features; A decoder is used to decode the multi-scale difference features to obtain a change probability map.

[0068] In specific implementation, the above modules and / or units can be implemented as independent entities, or they can be arbitrarily combined and implemented as the same or several entities. For the specific implementation of the above modules and / or units, please refer to the previous method embodiments. For the specific beneficial effects that can be achieved, please also refer to the beneficial effects in the previous method embodiments, which will not be repeated here.

[0069] The foregoing has provided a detailed description of a remote sensing image change detection method and system based on Fourier frequency domain feature enhancement provided by the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting changes in remote sensing images based on Fourier frequency domain feature enhancement, characterized in that, The method includes: Acquire dual-temporal images: the first image and the second image; The first image and the second image are passed through a shallow convolutional stream with shared weights to obtain a first initial feature map and a second initial feature map; The first initial feature map and the second initial feature map are respectively input into the corresponding multi-cascade encoder to obtain the first deep feature and the second deep feature; The first deep feature and the second deep feature are input into the adaptive dual-temporal interaction module to perform cross-temporal interaction and guided fusion, generating multi-scale difference features; The multi-scale difference features are input into the decoder to obtain the change probability map.

2. The remote sensing image change detection method based on Fourier frequency domain feature enhancement according to claim 1, characterized in that, The two multi-cascade encoders have the same structure, both including multiple frequency domain feature enhancement and hybrid extraction modules connected in sequence. Each frequency domain feature enhancement and hybrid extraction module includes a local feature path and a global frequency domain feature path. The processing procedure of the multi-cascade encoder includes: After performing layer normalization on the initial feature map, a normalized initial feature map is obtained. The normalized initial feature map is then input into the local feature path and the global frequency domain feature path to obtain local features and global frequency domain features, respectively. The local features and the global frequency domain features are fused and then linearly mapped to obtain the mapped features. The mapped features and the initial feature map are then merged through scale residual connection to obtain the deep features.

3. The remote sensing image change detection method based on Fourier frequency domain feature enhancement according to claim 2, characterized in that, The processing of the local feature path includes: The normalized initial feature map is passed through a two-dimensional convolutional layer and then through an activation function to obtain local features.

4. The remote sensing image change detection method based on Fourier frequency domain feature enhancement according to claim 2, characterized in that, The processing procedure for the global frequency domain feature path includes: Flatten the normalized initial feature map to obtain the feature sequence; Perform a discrete Fourier transform on the feature sequence to obtain a complex frequency spectrum, and decompose the complex frequency spectrum into a real part and an imaginary part; The real and imaginary parts are recombined, and the spatiotemporal sequence is reconstructed back using inverse Fourier transform; The spatiotemporal sequence is linearly mapped to obtain global frequency domain features.

5. The remote sensing image change detection method based on Fourier frequency domain feature enhancement according to claim 1, characterized in that, The adaptive dual-temporal interaction module includes parallel sub-modules: a long-range context interaction unit and a fine spatial alignment module. The processing procedure of the adaptive dual-temporal interaction module includes: Using the second deep feature as the guiding feature and the first deep feature as the guided feature, the first deep feature and the second deep feature are input into the long-range context interaction unit of the submodule to obtain the first global guiding feature. The first deep feature and the second deep feature are input into the fine spatial alignment module to obtain the first local spatial feature. Using the first deep feature as the guiding feature and the second deep feature as the guided feature, the first deep feature and the second deep feature are input into the long-range context interaction unit of the submodule to obtain the second global guiding feature. The first deep feature and the second deep feature are input into the fine spatial alignment module to obtain the second local spatial feature. Based on a dynamic gating mechanism, the first global guidance feature and the first local spatial feature are fused to obtain a first fused feature, and the second global guidance feature and the second local spatial feature are fused to obtain a second fused feature; Multi-scale difference features are obtained based on the first fusion feature and the second fusion feature.

6. The remote sensing image change detection method based on Fourier frequency domain feature enhancement according to claim 5, characterized in that, Using the second deep feature as the guiding feature and the first deep feature as the guided feature, the first and second deep features are input into the long-range context interaction unit of the submodule to obtain the first global guiding feature. The first and second deep features are then input into the fine spatial alignment module to obtain the first local spatial feature, including: The first deep feature is obtained by applying a linear transformation and an activation function to the first feature. The first deep feature is input into the selective state space model after linear transformation and convolutional layer to obtain the first context information. The first context information is then normalized by layer to obtain the normalized first context information. The second feature is obtained by fusing the first feature and the normalized first context information; The second deep feature is processed through a linear transformation and a convolutional layer and then input into a selective state space model to obtain the corresponding second context information. The second context information is then processed through an activation function to obtain the transformed second context information. The second feature, the normalized first context information, and the transformed second context information are fused to obtain the first global guiding feature; The third feature is obtained by applying a linear transformation and activation function to the first deep feature; The first deep feature is input into the selective state space model after linear transformation and convolutional layer to obtain the third context information. The third context information is then normalized by layer to obtain the normalized third context information. The fourth feature is obtained by fusing the third feature and the normalized third context information; The second deep feature is passed sequentially through a convolutional layer, an activation function, another convolutional layer, and another activation function to obtain the fifth feature; The fourth feature, the normalized third contextual information, and the fifth feature are fused to obtain the first local spatial feature.

7. The remote sensing image change detection method based on Fourier frequency domain feature enhancement according to claim 5, characterized in that, Based on a dynamic gating mechanism, the first global guidance feature and the first local spatial feature are fused to obtain a first fused feature, and the second global guidance feature and the second local spatial feature are fused to obtain a second fused feature, including: The first global guiding feature and the first local spatial feature are compressed along the spatial dimension and then concatenated. Then, through a linear layer and an activation function, two sets of first gating scores are obtained. The first global guidance feature and the first local spatial feature are weighted and fused using the two sets of first gating scores to obtain the first fused feature; The second global guiding feature and the second local spatial feature are compressed along the spatial dimension and then concatenated. Then, through a linear layer and an activation function, two sets of second gating scores are obtained. The second global guiding feature and the second local spatial feature are weighted and fused using the set of second gating scores to obtain the second fused feature.

8. The remote sensing image change detection method based on Fourier frequency domain feature enhancement according to claim 7, characterized in that, The formula for calculating the first gate score for both groups is: in, The first gate score, This indicates global guidance features. Representing local spatial features, For activation function, For linear layers, For splicing operations, This indicates compression along the spatial dimension. The above equation yields two sets of normalized first-gated scores. and .

9. The remote sensing image change detection method based on Fourier frequency domain feature enhancement according to claim 7, characterized in that, The first global guidance feature and the first local spatial feature are weighted and fused using the two sets of first gating scores to obtain the first fused feature, which is expressed by the following formula: in, For element-wise multiplication, This is the first fusion feature.

10. A remote sensing image change detection system based on Fourier frequency domain feature enhancement, characterized in that, include: The acquisition module is used to acquire dual-temporal images: a first image and a second image; The feature extraction module is used to pass the first image and the second image through a shallow convolutional stream with shared weights to obtain a first initial feature map and a second initial feature map; A multi-cascade encoder is used to input the first initial feature map and the second initial feature map into the corresponding multi-cascade encoder to obtain the first deep feature and the second deep feature. An adaptive dual-temporal interaction module is used to perform cross-temporal interaction and guided fusion of the first deep feature and the second deep feature to generate multi-scale difference features; A decoder is used to decode the multi-scale difference features to obtain a change probability map.

Images