Lightweight sensing system for complex environment change and detection capability enhancing method

Abstract

The invention discloses a lightweight sensing system for complex environment change. The lightweight sensing system comprises an edge enhancement module, a MobileNetV3 backbone network, a deformable feature alignment module, a feature difference calculation layer and a feature fusion classification module which are sequentially connected through signals. The invention further discloses a complex environment change detection capability enhancing method, a balanced lightweight design and high-precision deep learning remote sensing change detection model is provided through the complex environment change lightweight sensing system, and high-precision change area identification in a complex environment is realized. The lightweight model has the potential of being deployed on edge terminals with limited computing resources such as mobile equipment to run in real time, meanwhile, the robustness is high, interference such as illumination and shadow can be resisted, detail edges can be accurately extracted, and the capacity of adapting to space deformation is achieved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image analysis technology, and relates to a lightweight sensing system for complex environmental changes. This invention also relates to a method for enhancing the detection capability of complex environmental changes. Background Technology

[0002] With the rapid development of remote sensing technology, remote sensing image analysis has been widely applied in environmental monitoring, land use, UAV navigation, disaster assessment, and other fields. Remote sensing image change perception detection aims to identify areas of change on the surface of complex environments by comparing and analyzing remote sensing images from different time points, which has significant practical implications. However, traditional change detection methods (such as interpolation and thresholding methods) are easily affected by factors such as lighting and climate, leading to insufficient detection accuracy. In recent years, deep learning technology has demonstrated excellent performance in image recognition and feature extraction, providing new solutions for change detection. Current mainstream solutions are mostly based on convolutional neural networks (CNNs), which, while improving detection accuracy, still have the following problems: insufficient feature representation, making it difficult to capture complex changes; high computational cost and number of parameters, making it difficult to deploy on mobile devices; insufficient adaptability of the model to changes at different scales; and high sensitivity to spatial deformation during change matching.

[0003] Some methods currently employ the MobileNet series as a backbone network to improve detection speed, but they suffer from the following technical problems: low detection accuracy, as traditional networks use fixed sampling or simple upsampling to align multi-scale features, which cannot adapt to deformations in the input image (such as translation, scaling, and distortion), leading to feature "mismatch"; insufficient information utilization, as multi-scale feature fusion only uses stitching or simple addition without combining "feature difference calculation," failing to highlight key difference information; and loss of detailed features, lacking targeted edge enhancement mechanisms, resulting in weak ability to capture subtle edges and small target features in the image, affecting subsequent classification or detection accuracy. Currently, existing remote sensing image change detection methods mostly employ traditional difference methods, image registration combined with threshold segmentation, or simple convolutional neural networks such as CNN, UNet, and FCN. Deep neural network architectures based on cross-temporal Transformers, such as BiT_CD and STADE-CDNet, are also effective. The former uses the Transformer as its core, while the latter introduces a difference enhancement module to extract important features from the difference map to detect changing regions, and a temporal memory module to extract temporal and spatial information. These architectures achieve relatively accurate change detection, but they also come with significant computational costs and a large number of parameters. Lightweight network architectures, such as LCD-Net and LHDACT, are another option. LCD-Net uses MobileNetV2 as its encoder, extracts features from dual-temporal images, and enhances the correlation between dual-temporal features through temporal interaction and fusion modules, thereby improving temporal context awareness. LHDACT is a lightweight hybrid dual-attention CNN and TR architecture network based on deep overparameterized convolutions (DO-Conv). Although both achieve model lightweighting, there is still room for improvement in their accuracy.

[0004] The algorithms described above have expanded the understanding and analysis of changing scenes in remote sensing images. However, the limitations of these methods restrict their final performance in complex scenes. Therefore, there is an urgent need for a lightweight deep learning model that combines multi-scale feature fusion, edge enhancement, and spatial deformation adaptation capabilities to achieve a balance between lightweight design and high-precision change detection in remote sensing images. Summary of the Invention

[0005] The first objective of this invention is to provide a lightweight sensing system for complex environmental changes, which solves the problems of weak ability to capture changes and large number of model parameters and computational overhead in the prior art.

[0006] A second objective of this invention is to provide a method for enhancing the detection capability of complex environmental changes.

[0007] The first technical solution adopted in this invention is a lightweight perception system for complex environmental changes, including an edge enhancement module, a MobileNetV3 backbone network, a deformable feature alignment module, a feature difference calculation layer, and a feature fusion classification module connected sequentially by signals.

[0008] The invention is further characterized by: The edge enhancement module dynamically enhances the edge features of dual-temporal remote sensing images at different times using a learnable Sobel operator; The core modules of the MobileNetV3 backbone network include inverse residual blocks, depthwise separable convolutions, SE attention modules, and the activation function h-swish; these are used to transform the bi-temporal enhanced features into three different scales of bi-temporal feature maps. The deformable feature alignment module includes an offset prediction branch, a mask prediction branch, a deformable alignment operation, and a residual connection. Each pair of bi-temporal feature maps of the same scale obtained through the MobileNetV3 backbone network is input into the deformable feature alignment module, and the aligned features are output. The feature difference calculation layer includes element-wise subtraction and unified dimension; by calculating the difference on the aligned multi-scale features, the dual-temporal features are mapped to the change response, the input is the features before change and the aligned features, and the output is the difference features; The feature fusion classification module includes a feature fusion layer and a classification layer; it fuses, reconstructs, and maps multi-scale differential features into the final prediction output.

[0009] The second technical solution adopted in this invention is a method for enhancing the detection capability of complex environmental changes. This method utilizes the aforementioned lightweight sensing system for complex environmental changes to enhance the detection capability of complex situation changes, specifically including the following steps: Step 1: Collect dual-temporal remote sensing images and ; Step 2: Process the dual-temporal remote sensing images and Edge enhancement is performed to obtain an enhanced image. and ; Step 3: For each and Extracting multi-scale feature maps and ; Step 4: Convert the multi-scale feature map and Perform adaptive alignment and output enhanced spatial alignment features. ; Step 5: Calculate the difference features at each scale. ; Step 6: Analyze the differences in characteristics. The output channels and scales are kept consistent, and then a change map is obtained through a classifier.

[0010] The invention is further characterized by: The specific steps for step 1 are as follows: Step 1-1: Preprocess the dual-temporal optical remote sensing images to ensure that the geometric and radiometric characteristics of the two temporal data remain consistent. Steps 1-2: Label the changed parts in the dual-temporal optical remote sensing images to obtain the corresponding change label maps; Steps 1-3: Crop the preprocessed dual-temporal optical remote sensing images and their corresponding change label maps to a uniform size to obtain dual-temporal remote sensing images. and The training samples.

[0011] The specific steps for step 2 are as follows: Step 2-1: Apply two convolutional layers to the input dual-temporal remote sensing image. , Feature extraction , ; in, (1), In formula (1), This represents the convolution operation. This is the output 32-channel feature map; Step 2-2: Use the learnable Sobel operator to extract features from the dual-temporal remote sensing images. , Calculate the horizontal gradient separately , and vertical gradient , ; in, (2), (3), In formulas (2) and (3), Indicates the convolution operation; For the initialization of the Sobel-x core, The Sobel-y kernel is initialized; Steps 2-3, will and The two kernels are located in the feature map. , Slide upwards and perform convolution calculations to obtain each pixel in the two gradient component maps. gradient value and ; Steps 2-4: Combine the two gradient values ​​corresponding to each pixel to obtain the gradient magnitude of each point. and ; in, (6), In formula (6) For feature map gradient magnitude, For feature map The gradient magnitude; Steps 2-5: Convolution to convert edge amplitudes , Mapping back to the input channel dimension yields edge features. , ; in, (7), In formula (7) Indicates the convolution operation; Steps 2-6: Overlay the edge features with the original input to obtain the enhanced dual-temporal image. and ; in, (8); Pixels in steps 2-3 and The calculation process is as follows: (4), (5).

[0012] In step 3 For phase-time one multi-scale feature maps, including , , ; For phase-time two-scale feature maps, including , , .

[0013] The specific steps of step 4 are as follows: Step 4-1: Given a phase-time feature of the input and misaligned phase-time features Predicted offset and weight mask ; in, (9), (10) In formula (9) This represents the offset of each sampling point ( Directions, a total of 3×3×2=18 channels); in formula (10) This represents the weight of each sampling point (3×3=9 channels), after Sigmoid normalization. ; Step 4-2: Utilize the predicted offset and weight mask ,right Perform deformable convolution calculations to obtain aligned features. : (11), in, Spatial coordinates These are fixed sampling points for a standard 3×3 convolution kernel. It is a location First Predicted offset of each sampling point It is a location First The weight of each sampling point Learnable convolutional kernel weights; Step 4-3, Aligned Features Enhanced aligned features are obtained through residual connections. ; in, (12).

[0014] The calculation in step 5 uses element-wise subtraction to calculate the phase-time feature at each scale. Features aligned with enhancements The differences were analyzed to obtain the difference features at each scale. ; in, (13).

[0015] The change diagram is shown as follows : (14) in, Indicates an upsampling operation; , , Indicates the convolution operation; For classifiers.

[0016] The beneficial effects of this invention are: This invention provides a method for enhancing the detection capability of complex environmental changes. Through a lightweight sensing system for complex environmental changes, it offers a deep learning remote sensing change detection model that balances lightweight design with high precision. This model achieves high-precision identification of change areas in complex environments. The lightweight model has the potential to run in real time on edge terminals with limited computing resources, such as mobile devices. It is also robust, able to resist interference from lighting and shadows, and can accurately extract detailed edges, and has the ability to adapt to spatial deformation. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the model framework of the lightweight sensing system for complex environmental changes of the present invention. Figure 2 This is a schematic diagram of the feature alignment step in the method for enhancing the detection capability of complex environmental changes in this invention; Figure 3 This is a diagram showing the change detection results of Embodiment 1 of the present invention on the CDD dataset; Figure 4 This is a diagram showing the change detection results of Embodiment 2 of the present invention on the LEVIR-CD dataset; Figure 5 This is a diagram showing the change detection results on the SYSU-CD dataset in Embodiment 3 of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0019] This invention provides a lightweight perception system for complex environmental changes. Specifically, the goal is to design a lightweight model structure based on MobileNetV3 deep feature extraction, which integrates edge enhancement, multi-scale feature fusion, and deformable feature alignment. This enables a dual-temporal image change detection model based on deep learning, particularly MobileNetV3, edge enhancement, and deformable alignment.

[0020] This invention provides a lightweight sensing system for complex environmental changes, such as... Figure 1 As shown, it includes an edge enhancement module, a MobileNetV3 backbone network, a deformable feature alignment module, a feature difference calculation layer, and a feature fusion classification module, which are connected sequentially by signals.

[0021] The edge enhancement module dynamically enhances the edge features of dual-temporal remote sensing images at different times through a learnable Sobel operator, solving the problem of traditional edge detection being sensitive to changes in illumination and noise interference in complex environments. At the same time, it preserves the original information through residual connections, avoiding the loss of details caused by excessive edge enhancement.

[0022] The dual-temporal remote sensing images are input into the edge enhancement module. This module first extracts features from the dual-temporal remote sensing images through a feature extraction layer. Then, it uses a learnable Sobel operator to dynamically calculate the horizontal and vertical gradients of the feature map, fuses the horizontal and vertical gradients to generate an edge intensity map. Compared with the traditional fixed Sobel kernel, the learnable Sobel operator adapts to the edge direction sensitivity of different scenes. The edge map is mapped back to the channel dimension of the input image through an edge feature fusion layer to achieve channel matching of residual connections. Finally, residual connections are used to superimpose the original image and edge features to output the enhanced features, which retains low-frequency information while highlighting high-frequency edges and avoids gradient vanishing or over-enhancement.

[0023] The MobileNetV3 backbone network is a lightweight backbone network used to extract deep semantic features from remote sensing images, balancing computational resources and accuracy, and adapting to the low power consumption requirements of edge devices. Specifically, the features enhanced by the edge enhancement module are input into the MobileNetV3 backbone network, and three different scales of dual-temporal feature maps are output.

[0024] The core modules of the MobileNetV3 backbone network include inverse residual blocks, depthwise separable convolutions, SE attention modules, and the h-swish activation function. Inverse residual blocks efficiently extract multi-scale features through depthwise separable convolutions and the SE module, balancing receptive field and computational cost. Depthwise separable convolutions split standard convolutions into depthwise convolutions (spatial filtering) and pointwise convolutions (channel blending) to reduce the number of parameters. The SE attention module dynamically recalibrates channel importance, enhancing useful features (such as edges) and suppressing noise. The h-swish activation function replaces the traditional Swish, making computation more efficient.

[0025] The deformable feature alignment module dynamically aligns the local deformation features of dual-temporal images through deformable convolution, solving the misalignment problem caused by viewpoint differences and surface deformation. The module takes the dual-temporal feature maps of the same scale obtained through the MobileNetV3 backbone network as input to the deformable feature alignment module and outputs the aligned features.

[0026] The deformable feature alignment module includes an offset prediction branch, a mask prediction branch, a deformable alignment operation, and residual connections. The offset prediction branch generates an 18-channel offset for the original features using a 3×3 convolutional layer, predicting the local deformation offset of the changed features relative to the original features, guiding dynamic sampling position adjustments. The mask prediction branch outputs a 9-channel mask for the original features using another 3×3 convolutional layer, suppressing invalid deformation regions or noise and enhancing alignment robustness. The deformable alignment operation uses the offset and mask obtained from the original features to perform deformable convolution on the changed features, dynamically adjusting the sampling grid to achieve feature space alignment, resulting in aligned features. The residual connections add the aligned features to the original features, preserving original information while enhancing alignment differences.

[0027] The feature difference calculation layer maps the bi-temporal features to the change response by calculating the difference on the aligned multi-scale features. The input is the features before the change and the aligned features, and the output is the difference features.

[0028] The feature difference calculation layer includes element-wise subtraction and unified dimension. Element-wise subtraction is used to calculate the difference features corresponding to the same scale for each group; unified dimension is used to upsample the shallow-scale difference features and then map the channels to 32 through a 1×1 convolution conv1; medium-scale difference features and deep-scale difference features are mapped to 32 channels through convolution respectively.

[0029] The feature fusion classification module fuses, reconstructs, and maps multi-scale differential features into the final prediction output.

[0030] The feature fusion classification module consists of a feature fusion layer and a classification layer. The feature fusion layer fuses multi-scale differential features; the classification layer further integrates the fused intermediate features and maps them to the final prediction channels. It only uses two convolutional layers, the BatchNorm2d function, and the ReLU activation function, resulting in relatively small computation and parameter counts.

[0031] This invention also provides a method for enhancing the detection capability of complex environmental changes. Using the aforementioned lightweight sensing system for complex environmental changes, the detection capability for changes in complex situations is enhanced, such as... Figure 1 As shown, please follow these steps: Step 1: Collect dual-temporal remote sensing images and .

[0032] The specific steps are as follows: Step 1-1: Preprocess the dual-temporal optical remote sensing images to ensure that the geometric and radiometric characteristics of the two temporal data are consistent; preprocessing includes geometric registration, radiometric correction and image normalization; Steps 1-2: Label the changed parts in the dual-temporal optical remote sensing images to obtain the corresponding change label maps; Steps 1-3: Crop the preprocessed dual-temporal optical remote sensing images and their corresponding change label maps to a uniform size to obtain dual-temporal remote sensing images. and The training samples.

[0033] Step 2: Process the dual-temporal remote sensing images and Edge enhancement is performed to obtain the enhanced dual-temporal image. and .

[0034] The specific steps are as follows: Step 2-1: Apply two convolutional layers to the input dual-temporal remote sensing image. , Feature extraction , ; in, (1), In formula (1), This represents the convolution operation. This is the output 32-channel feature map.

[0035] Step 2-2: Use the learnable Sobel operator to extract features from the dual-temporal remote sensing images. , Calculate the horizontal gradient separately , and vertical gradient , ; in, (2), (3), In formulas (2) and (3), Indicates the convolution operation; For the initialization of the Sobel-x core, The Sobel-y kernel is initialized; Steps 2-3: Extract edge gradient information of the feature map from the horizontal and vertical orthogonal dimensions using the convolutional kernel weights of each channel; that is... and The two kernels are located in the feature map. , Slide upwards and perform convolution calculations to obtain each pixel in the two gradient component maps. gradient value and ; For each pixel in the image The calculation process is as follows: (4), (5).

[0036] Steps 2-4: Combine the two gradient values ​​corresponding to each pixel to obtain the gradient magnitude of each point. and ; The gradient magnitude represents the edge strength at that point, and the gradient magnitude uses the Euclidean distance. (6), in, For feature map gradient magnitude, For feature map The gradient magnitude.

[0037] Steps 2-5: Convolution to convert edge amplitudes , Mapping back to the input channel dimension yields edge features. , ; in, (7), in, This indicates a convolution operation.

[0038] Steps 2-6: Overlay the edge features with the original input to obtain the enhanced dual-temporal image. and ; in, (8).

[0039] Step 3: Process the edge-enhanced dual-temporal images separately. and Extracting multi-scale feature maps and .

[0040] For phase-time one multi-scale feature maps, including , , ; For phase-time two-scale feature maps, including , , Multi-scale feature map extraction was performed using the MobileNetV3 backbone network.

[0041] Step 4: Apply the phase-time dual features obtained in Step 3. Phase and time characteristics Perform adaptive alignment and output enhanced spatially aligned features. .

[0042] like Figure 2 As shown, the specific steps are as follows: Step 4-1: Given a phase-time feature of the input and misaligned phase-time features Predicted offset and weight mask ; in, (9), (10) In formula (9) This represents the offset of each sampling point ( Directions, a total of 3×3×2=18 channels); in formula (10) This represents the weight of each sampling point (3×3=9 channels), after Sigmoid normalization. .

[0043] Step 4-2: Utilize the predicted offset and weight mask ,right Perform deformable convolution calculations to obtain aligned features. : (11), in, Spatial coordinates These are fixed sampling points for a standard 3×3 convolution kernel. It is a location First Predicted offset of each sampling point It is a location First The weight of each sampling point These are learnable convolutional kernel weights.

[0044] Step 4-3, Aligned Features Enhanced aligned features are obtained through residual connections. ; in, (12).

[0045] Step 5: Calculate the difference features at each scale. .

[0046] Element-wise subtraction is used to calculate the phase-time feature at each scale. Features aligned with enhancements The differences were analyzed to obtain the difference features at each scale. ; in, (13).

[0047] Step 6: Differential features at each scale Upsampling and convolution are performed to keep the output channels and scale consistent. Finally, a classifier is used to obtain the transformation map. .

[0048] It can be represented as: (14) in, Indicates an upsampling operation; , , Indicates the convolution operation; For classifiers.

[0049] By implementing methods to enhance the detection capabilities of complex environmental changes, we can obtain: 1. Improved detection accuracy: As shown in Examples 4-6 and Comparative Examples 1-3, this method was evaluated on the CDD, LEVIR-CD and SYSU-CD datasets, with F1 scores of 95.28%, 93.30% and 85.72%, respectively, indicating that the method has good detection and localization accuracy. 2. Enhanced Detail Capture: As can be seen from the visualization results of Examples 4-6, the edge enhancement module significantly improves the detection rate of subtle differences in remote sensing images; 3. Strong compatibility: The backbone network is MobileNetV3, which balances accuracy and lightweight design. With only 3M parameters, it can be deployed on mobile devices and is suitable for real-time target detection applications in remote sensing images.

[0050] Example 1 In this embodiment, a lightweight perception system for complex environmental changes is used to enhance the detection capability of complex situation changes; the lightweight perception system for complex environmental changes includes an edge enhancement module, a MobileNetV3 backbone network, a deformable feature alignment module, a feature difference calculation layer, and a feature fusion classification module, which are connected in sequence by signals.

[0051] The method for enhancing the detection capability of complex environment changes in this embodiment uses the aforementioned lightweight sensing system for complex environment changes to enhance the detection capability of complex situation changes. Specifically, it is implemented according to the following steps: Step 1: Collect dual-temporal remote sensing images and ; The specific steps for step 1 are as follows: Step 1-1: Preprocess the dual-temporal optical remote sensing images to ensure that the geometric and radiometric characteristics of the two temporal data remain consistent. Steps 1-2: Label the changed parts in the dual-temporal optical remote sensing images to obtain the corresponding change label maps; Steps 1-3: Crop the preprocessed dual-temporal optical remote sensing images and their corresponding change label maps to a uniform size to obtain dual-temporal remote sensing images. and The training samples.

[0052] Step 2: Process the dual-temporal remote sensing images and Edge enhancement is performed to obtain an enhanced image. and ; Step 3: For each and Extracting multi-scale feature maps and ; Step 4: Convert the multi-scale feature map and Perform adaptive alignment and output enhanced spatial alignment features. ; Step 5: Calculate the difference features at each scale. ; Step 6: Analyze the differences in characteristics. The output channels and scales are kept consistent, and then a change map is obtained through a classifier.

[0053] Example 2 The method for enhancing the detection capability of complex environmental changes in this embodiment uses a lightweight sensing system for complex environmental changes to enhance the detection capability of complex situation changes. Specifically, it is implemented according to the following steps: Step 1: Collect dual-temporal remote sensing images and .

[0054] Step 2: Process the dual-temporal remote sensing images and Edge enhancement is performed to obtain an enhanced image. and ; The specific steps for step 2 are as follows: Step 2-1: Apply two convolutional layers to the input dual-temporal remote sensing image. , Feature extraction , ; in, (1), In formula (1), This represents the convolution operation. This is the output 32-channel feature map; Step 2-2: Use the learnable Sobel operator to extract features from the dual-temporal remote sensing images. , Calculate the horizontal gradient separately , and vertical gradient , ; in, (2), (3), In formulas (2) and (3), Indicates the convolution operation; For the initialization of the Sobel-x core, The Sobel-y kernel is initialized; Steps 2-3, will and The two kernels are located in the feature map. , Slide upwards and perform convolution calculations to obtain each pixel in the two gradient component maps. gradient value and ; Steps 2-4: Combine the two gradient values ​​corresponding to each pixel to obtain the gradient magnitude of each point. and ; in, (6), In formula (6) For feature map gradient magnitude, For feature map The gradient magnitude; Steps 2-5: Convolution to convert edge amplitudes , Mapping back to the input channel dimension yields edge features. , ; in, (7), In formula (7) Indicates the convolution operation; Steps 2-6: Overlay the edge features with the original input to obtain the enhanced dual-temporal image. and ; in, (8).

[0055] Step 3: For each and Extracting multi-scale feature maps and .

[0056] Step 4: Convert the multi-scale feature map and Perform adaptive alignment and output enhanced spatial alignment features. ; The specific steps of step 4 are as follows: Step 4-1: Given a phase-time feature of the input and misaligned phase-time features Predicted offset and weight mask ; in, (9), (10) In formula (9) This represents the offset of each sampling point ( Directions, a total of 3×3×2=18 channels); in formula (10) This represents the weight of each sampling point (3×3=9 channels), after Sigmoid normalization. .

[0057] Step 4-2: Utilize the predicted offset and weight mask ,right Perform deformable convolution calculations to obtain aligned features. : (11), in, Spatial coordinates These are fixed sampling points for a standard 3×3 convolution kernel. It is a location First Predicted offset of each sampling point It is a location First The weight of each sampling point These are learnable convolutional kernel weights.

[0058] Step 4-3, Aligned Features Enhanced aligned features are obtained through residual connections. ; in, (12).

[0059] Step 5: Calculate the difference features at each scale. ; The calculation in step 5 uses element-wise subtraction to calculate the phase-time feature at each scale. Features aligned with enhancements The differences were analyzed to obtain the difference features at each scale. ; in, (13).

[0060] Step 6: Analyze the differences in characteristics. The output channels and scales are kept consistent, and then a change map is obtained through a classifier.

[0061] Example 3 The method for enhancing the detection capability of complex environmental changes in this embodiment uses a lightweight sensing system for complex environmental changes to enhance the detection capability of complex situation changes. Specifically, it is implemented according to the following steps: Step 1: Collect dual-temporal remote sensing images and ; Step 2: Process the dual-temporal remote sensing images and Edge enhancement is performed to obtain an enhanced image. and ; Step 3: For each and Extracting multi-scale feature maps and ; In step 3 For phase-time one multi-scale feature maps, including , , ; For phase-time two-scale feature maps, including , , .

[0062] Step 4: Convert the multi-scale feature map and Perform adaptive alignment and output enhanced spatial alignment features. ; Step 5: Calculate the difference features at each scale. ; Step 6: Analyze the differences in characteristics. The output channels and scales are kept consistent, and then a change map is obtained through a classifier.

[0063] Change diagram Represented as: (14) in, Indicates an upsampling operation; , , Indicates the convolution operation; For classifiers.

[0064] Example 4 In this embodiment, the evaluation is performed on the CDD dataset, using a method to enhance the detection capability of complex environmental changes. The specific steps are as follows: Step 1: Collect dual-temporal remote sensing images and ; Step 2: Process the dual-temporal remote sensing images and Edge enhancement is performed to obtain an enhanced image. and ; Step 3: For each and Extracting multi-scale feature maps and ; Step 4: Convert the multi-scale feature map and Perform adaptive alignment and output enhanced spatial alignment features. ; Step 5: Calculate the difference features at each scale. ; Step 6: Analyze the differences in characteristics. The output channels and scales are kept consistent, and then a change map is obtained through a classifier.

[0065] like Figure 3 As shown in the figure, the change detection results on the CDD dataset in this embodiment clearly demonstrate that this method can accurately extract the details and edges.

[0066] Example 5 This embodiment evaluates the method on the LEVIR-CD dataset, using a method to enhance the detection capability of complex environmental changes. The specific steps are as follows: Step 1: Collect dual-temporal remote sensing images and ; Step 2: Process the dual-temporal remote sensing images and Edge enhancement is performed to obtain an enhanced image. and ; Step 3: For each and Extracting multi-scale feature maps and ; Step 4: Convert the multi-scale feature map and Perform adaptive alignment and output enhanced spatial alignment features. ; Step 5: Calculate the difference features at each scale. ; Step 6: Analyze the differences in characteristics. The output channels and scales are kept consistent, and then a change map is obtained through a classifier.

[0067] like Figure 4 As shown in the figure, the change detection result on the LEVIR-CD dataset in this embodiment can achieve high-precision identification of change areas in complex environments.

[0068] Example 6 In this embodiment, the evaluation is performed on the SYSU-CD dataset, using a method to enhance the detection capability of complex environmental changes. The specific steps are as follows: Step 1: Collect dual-temporal remote sensing images and ; Step 2: Process the dual-temporal remote sensing images and Edge enhancement is performed to obtain an enhanced image. and ; Step 3: For each and Extracting multi-scale feature maps and ; Step 4: Convert the multi-scale feature map and Perform adaptive alignment and output enhanced spatial alignment features. ; Step 5: Calculate the difference features at each scale. ; Step 6: Analyze the differences in characteristics. The output channels and scales are kept consistent, and then a change map is obtained through a classifier.

[0069] like Figure 5 As shown in the figure, the change detection result on the SYSU-CD dataset in this embodiment can achieve high-precision identification of change areas in complex environments.

[0070] Comparative Example 1 In this comparative example, algorithms from different fields of remote sensing image change detection are evaluated on a CCD dataset. The evaluation results are shown in Table 1 below. In the table, MEF is the evaluation of the method for enhancing the detection capability of complex environmental changes in this application. BIT is a dual-temporal image transformer that models the global context of dual-temporal images by extracting semantic tokens, thereby improving detection accuracy while maintaining a lightweight model.

[0071] ChangeFormer is a twin network architecture based on Transformer, employing a hierarchical Transformer encoder and a lightweight MLP decoder. It can efficiently capture long-range details at multiple scales and performs excellently in remote sensing image change detection tasks.

[0072] CGNet is a Change Guiding Network. It uses change maps generated from deep features as prior information to guide the fusion of multi-scale features and designs a change guidance module to capture long-distance dependencies between pixels, thereby improving the expressive power of change features.

[0073] CF-GCN is a change detection method that combines graph convolutional networks (GCN) to model feature associations and spatial relationships in change detection.

[0074] STADE-CD leverages the features of the CD dataset, composed of spectral vectors in both temporal and spatial dimensions, to address the challenges of spurious changes and inherent imbalances in the RS change detection dataset. ChangeMamba is a remote sensing image change detection method based on the Mamba architecture. It applies Mamba to this field, utilizing its efficient spatiotemporal state-space model to process multi-temporal information, and is suitable for various tasks such as binary change detection and semantic change detection.

[0075] DDRL is a change detection method that combines deep reinforcement learning, optimizing the detection process through reinforcement learning mechanisms.

[0076] CDMamba is a change detection model that combines convolution and Mamba. It proposes a Scaled Residual ConvMamba module and an adaptive global-local guidance fusion module, which balances global modeling capabilities with local detail capture.

[0077] Table 1

[0078] As shown in Table 1, the F1 score is the balanced F-score, which is the harmonic mean of precision and recall. The method of the present invention achieved an F1 score of 95.28% on the CCD dataset, indicating that the method of the present invention for enhancing the detection capability of complex environmental changes has better detection and localization accuracy.

[0079] Comparative Example 2 In this comparative study, algorithms from different fields of remote sensing image change detection were evaluated on the LEVIR-CD dataset. The evaluation results are shown in Table 2 below. In the table, MEF is the method for enhancing the detection capability of complex environmental changes proposed in this application; SFEARNet is the Feature Enhancement and Adaptive Fusion Network.

[0080] Table 2

[0081] As shown in Table 2, the method of the present invention achieved an F1 score of 93.30% on the LEVIR-CD dataset, indicating that the method of the present invention for enhancing the detection capability of complex environmental changes has better detection and localization accuracy.

[0082] Comparative Example 3 In this comparative example, algorithms from different fields of remote sensing image change detection were evaluated on the SYSU-CD dataset. The evaluation results are shown in Table 3 below. In the table, MEF is the method for enhancing the detection capability of complex environmental changes in this application, and the other methods are algorithms used in the field of remote sensing image change detection in the prior art. Table 3

[0083] As shown in Table 3, the method of the present invention achieved an F1 score of 85.72% on the SYSU-CD dataset, indicating that the method of the present invention for enhancing the detection capability of complex environmental changes has good detection and localization accuracy.

Claims

1. A lightweight system for complex environmental change perception, characterized in that, The edge enhancement module, the MobileNetV3 backbone network, the deformable feature alignment module, the feature difference calculation layer and the feature fusion classification module are sequentially connected by signals.

2. The system, as claimed in claim 1, wherein, The edge enhancement module dynamically enhances the edge features of the double-time-phase remote sensing images of different periods through a learnable Sobel operator. The MobileNetV3 backbone network core module includes a reverse residual block, a depth separable convolution, an SE attention module and an activation function h-swish; and is used for converting the double-time-phase enhanced features to output double-time-phase feature maps of three different scales. The deformable feature alignment module includes an offset prediction branch, a mask prediction branch, a deformable alignment operation and a residual connection; each group of double-time-phase feature maps of the same scale obtained through the MobileNetV3 backbone network is input into the deformable feature alignment module, and aligned features are output. The feature difference calculation layer includes element-by-element subtraction and uniform dimension; the features of the double-time-phase are mapped to change responses by calculating the differences on the aligned multi-scale features, the input is the features before the change and the aligned features, and the output is the difference features. The feature fusion classification module includes a feature fusion layer and a classification layer; the multi-scale difference features are fused, reconstructed and mapped to the final prediction output.

3. A method for enhancing the detection capability of complex environmental changes, characterized by, The complex environment change lightweight perception system as claimed in claim 1 or 2 is used to enhance the detection capability of complex situation changes, and the following steps are implemented: Step 1, Collecting dual temporal remote sensing images and ; Step 2, edge enhancement on the dual temporal remote sensing image and to obtain an enhanced image and ; Step 3, respectively to and extract multi-scale feature maps and ; Step 4, performing adaptive alignment on the multi-scale feature maps and output enhanced spatially aligned features ; Step 5, calculating the difference features for each scale ; Step 6, Differential features The output channels and scales are made consistent by processing, and then the change map is obtained by the classifier.

4. The method of claim 3, wherein, The specific steps of step 1 are as follows: Step 1-1, pre-processing the double-time optical remote sensing images to ensure that the two-time-phase data are consistent in geometry and radiation characteristics; Step 1-2, labeling the changed part in the double-time optical remote sensing images to obtain the corresponding change label map; Step 1-3, the pre-processed dual-time optical remote sensing images and the corresponding change label maps are uniformly sized and cropped to obtain dual-time remote sensing images and training samples.

5. The method of claim 3, wherein, The specific steps of step 2 are as follows: Step 2-1, two-layer convolution on input dual-phase remote sensing images , extracting features , ; Wherein, （1）， In formula (1), denotes a convolution operation, is the output 32-channel feature map; Step 2-2, using a learnable Sobel operator to extract features of the two-phase remote sensing images , Calculate horizontal gradient , and vertical gradient , respectively; Wherein, （2）， （3）， In formulas (2) and (3), denotes a convolution operation; is an initialized sobel-x kernel, is an initialized sobel-y kernel; Step 2-3, the two kernels are respectively slid on the feature map and convolution calculation is performed to obtain gradient values of each pixel point in the two gradient component maps , . , ;​​​ Step 2-4, combine the two gradient values corresponding to each pixel point, respectively, to obtain the gradient amplitude of each point and ; Wherein, （6）， In Equation (6) is the gradient magnitude of the feature map is the gradient magnitude of the feature map is the gradient magnitude of the feature map is the gradient magnitude of the feature map Step 2-5, edge magnitude is mapped back to input channel dimension by convolution , , , ; Wherein, （7）， In equation (7) denotes a convolution operation; Step 2-6, superimpose edge features with original input to get enhanced dual-phase image and ; Wherein, （8）。 6. The method of claim 5, wherein, The pixel point in the step 2-3 And The calculation process is: （4）， （5）。 7. The method of claim 3, wherein, In step 3 is a multi-scale feature map for phase one, comprising , , ; is a multi-scale feature map for phase two, comprising , , .

8. The method of claim 3, wherein, The specific steps of step 4 are as follows: Step 4-1, given input phase time one feature and misaligned phase time two feature , predict offset and weight mask ; Wherein, （9）， （10）， In formula (9) represents the offset of each sampling point (3x3=9 channels) direction, a total of 3x3x2=18 channels); In formula (10) represents the weight of each sampling point (3x3=9 channels), which is normalized to by Sigmoid; Step 4-2, using the predicted offset and weight masks , to perform deformable convolution computation to get aligned features : （11）， wherein, is a spatial coordinate position, is a fixed sampling point of a standard 3x3 convolution kernel, is a prediction offset of a th sampling point at position is a weight of a th sampling point at position is a weight of a th sampling point at position is a learnable convolution kernel weight; Step 4 - 3, aligned features Through residual connections, get enhanced aligned features ; Wherein, （12）。 9. The method of claim 3, wherein, The calculation in step 5 employs element-wise subtraction, calculating the difference of each scale phase with a feature with enhanced aligned features to obtain the difference feature of each scale ; Wherein, （13）。 10. The method of claim 7, wherein, The change map is represented as : （14）， wherein, denotes an up-sampling operation; , , denotes a convolution operation; is a classifier.

Images