Change monitoring algorithm and system based on satellite image super-resolution

Abstract

The invention relates to an efficient super-resolution processing and accurate change monitoring algorithm and system based on satellite images. The system combines a blind image super-resolution model and a change monitoring model based on a Transform architecture, and aims to solve the problems of insufficient satellite image resolution and low change monitoring precision in the prior art. Experimental verification results show that the method has excellent performance in the aspects of super-resolution reconstruction and change monitoring, and the image resolution and the change monitoring precision can be remarkably improved. The system has wide application prospects in the fields of urban planning, environmental protection, disaster early warning and the like, and provides powerful technical support for related fields.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, especially for the super-resolution reconstruction and change detection of satellite images. By combining a bidirectional spatial residual generative adversarial network with a change detection model based on the Transformer architecture, a change detection technology for ultra-large satellite remote sensing images is realized. The present invention aims to achieve efficient super-resolution processing of satellite images and accurate surface change monitoring, providing strong technical support for fields such as urban planning, environmental protection, and disaster warning. Background Art

[0002] With the rapid development of satellite remote sensing technology, high-resolution satellite images have become an important means of earth observation. However, limited by hardware conditions and transmission bandwidth, there are still many challenges in directly obtaining high-resolution images. At the same time, surface change monitoring requires efficient and accurate analysis of time-series images, and traditional methods have limitations in dealing with complex scenarios and subtle changes. Therefore, it is particularly important to develop an algorithm that can effectively improve the image resolution and accurately monitor surface changes.

[0003] As an advanced image super-resolution technology, the blind image super-resolution model effectively improves the detail fidelity and spatial coherence of the generated images by introducing a bidirectional spatial residual learning mechanism. The CF change detection model, based on the Transformer architecture, is good at capturing long-range dependencies in images and is highly sensitive to surface changes. Combining the advantages of both, the present invention aims to achieve precise change detection through the efficient super-resolution reconstruction of satellite images. Summary of the Invention

[0004] 1. Technical Problem

[0005] Aiming at the problems of insufficient resolution of satellite images and low accuracy of change detection in the prior art, the present invention provides a satellite image super-resolution change detection algorithm and system based on the blind image super-resolution model and the CF change detection model. The system aims to achieve high-precision super-resolution reconstruction of images through the blind image super-resolution model, and use the CF change detection model to perform change detection on the reconstructed images, so as to achieve automatic and rapid identification of surface changes.

[0006] 2. Technical Solution

[0007] The technical solution of the present invention includes key steps such as data preprocessing, super-resolution reconstruction by the blind image super-resolution model, change detection by the CF change detection model, and system integration and optimization. The technical parameters and formulas of each step will be introduced in detail below.

[0008] (1) Data Preprocessing

[0009] The data preprocessing stage mainly includes image registration, noise removal, and image enhancement. Image registration is used to ensure the spatial consistency between time-series images and is achieved through feature matching and geometric transformation. Noise removal uses filtering algorithms to reduce noise interference in the images and improve the accuracy of subsequent processing. Image enhancement improves the visualization effect of the images through operations such as contrast adjustment and sharpening.

[0010] (2) Blind Image Super-Resolution Model Super-Resolution Reconstruction

[0011] In the super-resolution reconstruction stage of the blind image super-resolution model, a bidirectional spatial residual generative adversarial network is used for the super-resolution processing of images. This network consists of a generator G and a discriminator D. The generator G is responsible for generating high-resolution images, while the discriminator D is used to distinguish between real high-resolution images and generated images.

[0012] Its core ideas are divided into blind degradation, the RRDB main body, multi-scale confrontation, pixel fusion, perceptual and adversarial losses

[0013] Generator G Structure: The generator G uses bidirectional spatial residual blocks, combined with upsampling and downsampling paths, to enhance the image feature extraction and reconstruction capabilities. The bidirectional spatial residual block enables the network to capture both local details and global structural information of the images by introducing spatial residual connections.

[0014] The generator is denoted as G 0 , for the input low-resolution image I lr and the output high-resolution image I sr The result can be expressed as: Head Convolution: f 0 = W H * I lr + b H ; RRDBGAN: F rrdb = RRDB N ....RRDB 1 (F 0 ) where there are N RRDBs in series; Global Residual: F trunk = F 0 + W trunk * F rrdb ; Upsampling uses PixelShuffle to enlarge the feature map from the LR space to the HR space; Output Convolution I sr = W tail * F up + b tail

[0015] Discriminator D Structure: The discriminator D adopts a convolutional neural network structure and is used to distinguish between real high-resolution images and generated images. Through continuous training, the discriminator D can gradually improve its ability to judge the authenticity of generated images, thereby guiding the generator G to generate higher-quality images.

[0016] The discriminator often adopts a multi-scale discriminator or a relative discriminator. This algorithm uses a single-scale discriminator. Let D φ Output the score for the input image being real (where φ is a parameter)

[0017] For the real image I hr Give the score D φ (I hr ); For the generated image I SR Give the score D φ (I SR );

[0018] Loss Function: It includes pixel loss, perceptual loss, adversarial loss, and overall loss. The formula involved in pixel loss is: It helps the generator maintain a basic restoration degree in the initial stage, and generally has a small weight. The perceptual loss uses several layers of a pre-trained network such as vgg19 to extract features Φ 1 Then perform MSE / L1 on the feature layer: The discriminator in the adversarial loss is: For the generator: So the overall loss function formula is: Each coefficient needs to be continuously tuned in the experiment to finally obtain the optimal parameters.

[0019] Training Strategy: Use the Adam optimizer for training. The initial learning rate is set to 0.0002 and gradually decays as the training progresses. The batch size is adjusted according to the hardware conditions, and 16 or 32 is generally recommended. During the training process, the parameters of the generator G and the discriminator D are continuously updated through iteration until the convergence condition is reached.

[0020] (3) Change Detection

[0021] In the change detection stage, a change detection model based on the Transformer architecture is used to perform change detection on the super-resolution reconstructed images. This model can accurately identify surface changes by capturing long-range dependencies in the images.

[0022] Feature Extraction: Use a pre-trained CNN or Transformer model to extract feature maps from the super-resolution reconstructed images. The feature maps contain rich information of the images and are helpful for subsequent change detection tasks.

[0023] Change Detection: The feature maps of time-series images are used as inputs, and the self-attention mechanism in the change detection model is employed to capture the differences between images. The self-attention mechanism can calculate the correlation between any two positions in the feature map, thereby capturing the long-range dependencies in the images. The specific formula is as follows: where Q, K, and V are the query, key, and value matrices respectively, is the dimension of the key. The softmax function is used to normalize the scores so that the sum of the scores at each position is 1. By calculating the dot product of the query matrix Q and the key matrix K and dividing by for scaling, then applying the softmax function for normalization, and finally multiplying by the value matrix V to obtain the weighted sum as the output.

[0024] During change detection, the feature maps of time-series images are used as the query matrix Q and the key-value matrices K / V respectively, and the change feature map is obtained through the self-attention mechanism. Each pixel value in the change feature map indicates whether a change has occurred at the corresponding position. By setting a threshold for binary processing, a change mask is obtained.

[0025] Post-processing: Morphological operations (such as dilation and erosion) are performed on the change mask to optimize the boundaries of the changed areas and reduce false detections and missed detections. Through morphological operations, the accuracy and robustness of change detection can be further improved.

[0026] (4) System Integration and Optimization

[0027] Integrate the above steps into a unified software platform to support users in uploading satellite image data and automatically execute super-resolution reconstruction and change detection tasks. Meanwhile, introduce model lightweight technology to reduce computational resource consumption and improve processing speed. In addition, a user-friendly interface is provided to display the processing results and support result export and visualization analysis.

[0028] During system integration, a modular design is adopted, and the super-resolution reconstruction module of the blind image super-resolution model and the change detection module of the CF change detection model are encapsulated as independent components. Data interaction and collaborative work between modules are achieved through interface calls. Meanwhile, the system is optimized to improve processing efficiency and stability.

[0029] 3. Technical Parameters

[0030] The technical parameters of the present invention include network depth, learning rate, batch size, selection of feature extraction network, and configuration of the CF change detection model, etc. The specific parameter settings are as follows:

[0031] Network Depth: The number of bidirectional spatial residual blocks in the generator G of the blind image super-resolution model is set to N = 10 to balance the computational complexity and reconstruction quality.

[0032] Learning rate: The Adam optimizer is adopted, and the initial learning rate is set to 0.0002, which is gradually decayed as the training progresses. The decay strategy can adopt cosine annealing or step decay.

[0033] Batch size: Adjusted according to the hardware conditions, and generally recommended to be 16 or 32. The choice of batch size needs to balance memory occupancy and training efficiency.

[0034] Feature extraction network: Pretrained networks such as VGG, ResNet, or Transformer can be used as the feature extraction network, and the specific choice needs to be determined according to the experimental results.

[0035] CF change detection model configuration: The number of Transformer layers L is set to 6, the number of attention heads H is set to 8, and the embedding dimension D is set to 512. The selection of these parameters aims to balance the model's expressive ability and computational complexity, ensuring that the CF change detection model can efficiently capture long-range dependencies in the images and achieve accurate change detection.

[0036] 4. Experimental verification and result analysis

[0037] To verify the effectiveness of the present invention, a series of experiments were conducted. The experimental data was sourced from multiple satellite remote sensing datasets, including surface images with different resolutions and at different time points. Through comparative experiments, the performance of the present invention in super-resolution reconstruction and change detection was evaluated.

[0038] (1) Super-resolution reconstruction experiment

[0039] In the super-resolution reconstruction experiment, the blind image super-resolution model was compared with other advanced super-resolution algorithms (such as SRGAN, ESRGAN, etc.). The peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) were used as evaluation metrics. The experimental results showed that the blind image super-resolution model could generate richer detail information while maintaining the structural consistency of the images, and the PSNR and SSIM values were higher than those of the comparative algorithms.

[0040] (2) Change detection experiment

[0041] In the change detection experiment, the CF change detection model was compared with traditional change detection methods (such as differential image method, post-classification comparison method, etc.). Precision, Recall, and F1-score were used as evaluation metrics. The experimental results showed that the CF change detection model could more accurately identify surface changes, especially performing well in complex scenarios and subtle changes.

[0042] (3) System performance test

[0043] Performance tests were conducted on the integrated system, including aspects such as processing speed, resource occupancy, and stability. The test results show that the system can efficiently process large-scale satellite image data and exhibits good stability and scalability under different hardware configurations. Description of the Drawings

[0044] Figure 1 It is a schematic structural diagram of the present invention.

[0045] S1: Data preprocessing mainly includes image registration, noise removal, and image enhancement processing;

[0046] S2: They are the low-resolution pre-time image LR and the low-resolution post-time image LR after being processed by S1;

[0047] S3: Use the super-resolution model to perform super-resolution reconstruction on the pre-time image LR and the low-resolution post-time image LR corresponding to S2;

[0048] S4: They are the high-resolution pre-time image HR and the high-resolution post-time image HR data corresponding to the super-resolution reconstruction processing of S3;

[0049] S5: Use the Transformer model to extract feature maps from the high-resolution pre-time image HR and the high-resolution post-time image HR data corresponding to the super-resolution reconstruction processing of S4

[0050] S6: Use the trained change detection model CF to perform change detection extraction and generate change feature maps on the high-resolution pre-time image HR and the high-resolution post-time image HR data corresponding to the super-resolution reconstruction processing of S4 and S3

[0051] S7: Perform binarization processing on the change feature maps generated by S6 to generate a binary image with attributes only 0 and 1, then convert the binary image into a shapefile file to generate the vector range of the change detection patches, and then perform manual screening for each patch to modify the boundary of the shapefile range to make the monitoring range compatible with the actual change range;

[0052] S8: Final change detection results

[0053] 6. Conclusion

[0054] The present invention proposes a satellite image super-resolution change detection algorithm and system based on a blind image super-resolution model and a CF change detection model. By combining advanced super-resolution technology and a change detection model, efficient processing and accurate analysis of satellite images are achieved. Experimental results show that this method has achieved remarkable results in improving image resolution and capturing surface changes, providing strong technical support for fields such as urban planning, environmental protection, and disaster warning. In the future, the algorithm and system performance will be further optimized, application scenarios will be expanded, and the processing efficiency and accuracy will be improved.

Claims

1. A super-resolution change monitoring method based on satellite images, characterized in that: The following steps are involved: a) Data preprocessing step: registering, removing noise and enhancing the input satellite images; b) a blind image super-resolution model super-resolution reconstruction step, using a generative adversarial network to perform super-resolution processing on the pre-processed image to generate a high-resolution image, wherein the generative adversarial network includes a generator and a discriminator, and the generator adopts a bidirectional spatial residual block structure; c) a feature extraction step, extracting a feature map from the super-resolution reconstructed image; d) a change monitoring step, using a change monitoring model based on a Transformer architecture to perform change detection on the feature map of the time series image, generate a change feature map, and binarize the change feature map to obtain a change mask; e) Post-processing step, morphological operations are performed on the change mask to optimize the boundaries of the change area.

2. The method according to claim 1, characterized in that The bidirectional spatial residual block structure of the generator includes an upsampling path and a downsampling path to enhance the image feature extraction and reconstruction capabilities.

3. The method according to claims 1 and 2, characterized in that The loss functions of the generative adversarial network include pixel loss, perceptual loss and adversarial loss.

4. The method according to claim 1, characterized in that: The Transformer-based change monitoring model in the change monitoring step can accurately identify surface changes by capturing long-distance dependencies in images.

5. A satellite image super-resolution change monitoring system for implementing the method according to any one of claims 1 to 4, characterized in that: include: a) Data preprocessing module, used to perform registration, noise removal and image enhancement on the input satellite images; b) a blind image super-resolution model module, which is used to perform super-resolution processing on the pre-processed image to generate a high-resolution image; c) a feature extraction module, which is used to extract feature maps from the super-resolution reconstructed image; d) a change monitoring module, which uses a change monitoring model based on the Transformer architecture to perform change detection on the feature map of the time series image, generate a change feature map, and perform binarization processing on the change feature map to obtain a change mask; e) A post-processing module is used to perform morphological operations on the change mask and optimize the boundary of the change area.

Images