Land change detection method combining matrix decomposition and adaptive propagation, and system

Abstract

The present invention belongs to the technical field of land change detection. Disclosed are a land change detection method combining matrix decomposition and adaptive propagation, and a system. A deep learning semantic segmentation model combining matrix decomposition and adaptive propagation is used for detecting a land change. The model is of a twin double-branch structure, which can receive both an input of a pre-temporal remote sensing image data set and an input of a post-temporal remote sensing image data set and can share network parameters, and which has the advantage of end-to-end feature learning. In the present invention, an attention method based on matrix decomposition is used to replace a conventional attention method based on manual design; moreover, in view of adaptive propagation technology for pixel offsets of neighboring points, a deep learning twin differential semantic segmentation network is established to realize automatic and precise land utilization / land coverage change detection on the basis of high-resolution optical data.

Description

Land change detection method and system combining matrix decomposition and adaptive propagation Technical Field

[0001] The present invention belongs to the technical field of land change detection, and in particular relates to a land change detection method combining matrix decomposition and adaptive propagation. Background Art

[0002] Land use / land cover change detection uses remote sensing technology and change detection algorithms to quantitatively analyze surface changes in remote sensing imagery over different time periods over the same area. By detecting land change, we can understand a region's development trends, environmental changes, and the impact of human activities on the land. This is of great significance for sustainable development planning, resource management, and environmental protection. Traditional land change detection methods can be categorized as pixel-based or object-based, depending on the object unit. Pixel-based change detection methods only consider individual pixels and fail to account for spatial context. Object-based change detection methods, on the other hand, are superior in considering the spatial structure and context of the ground objects, but require more computational resources.

[0003] In recent years, deep learning-based land change detection methods have been widely used in remote sensing. Deep learning methods have achieved remarkable results in semantic segmentation tasks. Typical deep learning semantic segmentation networks typically employ a convolutional neural network architecture, extracting image feature information through multi-layer convolution and pooling operations. The feature maps are then restored to the input image resolution through upsampling or deconvolution, and classification is performed pixel by pixel. Convolutional neural network-based classification methods can achieve comparable results to traditional pixel- and object-based methods. Previous post-classification deep learning methods first train a convolutional neural network to classify remote sensing images at two different times, then compare the two classification results to extract change information. Compared to post-classification methods, twin semantic segmentation change detection methods achieve efficient end-to-end feature extraction. Furthermore, global attention mechanisms are widely used in semantic segmentation to improve the model's perceptual scope and semantic understanding capabilities. Global attention mechanisms allow the network to focus on important areas in the image, thereby improving the recognition accuracy of specific objects. However, global attention mechanisms have high computational and memory overhead, especially for large-scale, high-resolution images, which increases the complexity of training and inference. Therefore, finding a global attention algorithm with lower overhead and comparable performance is a worthy research topic. Furthermore, edge information is also crucial in semantic segmentation. To improve the ability to capture image details and structure, some methods have introduced edge optimization modules. Edge information helps to more accurately segment object edges, resulting in more refined and accurate segmentation results. Combining edge optimization with semantic segmentation to improve model performance and enhance the accuracy of land change detection is crucial. Technical issues

[0004] Through the above analysis, the problems and defects of the existing technology are as follows:

[0005] Traditional land change detection methods can be divided into two categories based on object units: pixel-based and object-based. Pixel-based change detection methods only consider a single pixel and do not consider spatial context. Object-based change detection methods are superior in considering the spatial structure and contextual information of land objects, but require more computing resources. Post-classification-based deep learning methods and single-stream networks cannot achieve end-to-end change detection. Twin semantic segmentation networks can share network parameters, reducing training time and resources. Semantic segmentation networks usually use global attention to enhance feature capture capabilities, but the computational and memory overhead is large. At the same time, semantic segmentation still has the problem of inaccurate edge segmentation. It is necessary to add a dedicated edge detection module to the network to enhance the model's edge perception ability. Technical Solutions

[0006] In view of the problems existing in the prior art, the present invention provides a land change detection method combining matrix decomposition and adaptive propagation.

[0007] The land change detection method of the present invention specifically uses a deep learning model that combines matrix decomposition and adaptive propagation to process and detect changes in high-resolution remote sensing data. The model uses a twin differential structure to achieve end-to-end learning of remote sensing images in the previous and next phases, and shares network parameters to improve feature learning efficiency. The key innovation is to introduce a global attention module based on matrix decomposition, abstract global information into a low-rank matrix, and learn through a neural network structure to reduce the computational and memory overhead of global information modeling and improve the computational efficiency of the model. At the same time, an adaptive propagation algorithm is used to solve the edge pixel sampling problem, and by learning fixed and variable two-dimensional offsets and weighted averaging local neighborhoods, the model's ability to capture the edges of objects is improved. The method of the present invention realizes automated and accurate land use / land cover change detection on high-resolution optical data.

[0008] Another object of the present invention is to provide a land change detection method combining matrix decomposition and adaptive propagation, comprising:

[0009] Step 1: Configure the deep learning environment;

[0010] Step 2: high-resolution change detection data acquisition and preprocessing;

[0011] Step 3: Build a land change detection model;

[0012] Step 4: Model training, verification and test evaluation.

[0013] Furthermore, the method for configuring the deep learning environment is as follows:

[0014] Based on the Pytorch open source deep learning framework, the operating environment is Ubuntu 20.04, Python version is 3.10, the processor is Intel(R) Core(TM) i9-10900X, the graphics card is NVIDIA GeForce RTX 3080, and the development environment is PyCharm.

[0015] Furthermore, the high-resolution change detection data acquisition and preprocessing method:

[0016] Two public datasets, WHU Building Change Detection and LEVIR-CD, were used;

[0017] There are a variety of buildings with large-scale changes in the WHU building change detection dataset;

[0018] The dataset consists of two cycles of aerial images and change labels. The two cycles were acquired in 2012 and 2016, respectively. The resolution of each image is 0.3 meters. Since there is no standard segmentation for this dataset, each image is cropped into 256×256 image slices and randomly divided into training / validation / test sets in a ratio of 7:1:2. There is no overlap in the slices during the allocation process.

[0019] The LEVIR-CD dataset consists of 637 ultra-high-resolution image pairs and corresponding change labels collected from Google Earth. The resolution of each image is 0.5 meters and the size is 1024×1024. It is a large-scale change detection dataset covering different types of buildings.

[0020] The LEVIR-CD dataset provides a standard train / validation / test split, which uses 70% of the samples for training, 10% for validation, and 20% for testing; similarly, the samples are cropped into image slices of size 256×256.

[0021] Furthermore, the land change detection model is constructed:

[0022] (1) Feature extraction backbone network

[0023] The model adopts a twin dual-branch structure with shared parameters, aiming to simultaneously learn the features of remote sensing images in the front and back phases. VGG-16 is used as the feature extraction backbone network, which consists of 16 convolutional layers and 3 fully connected layers. Multiple convolutional layers can learn multi-level feature representations. Since VGG-16 is large in model size and number of parameters, the results of the first three convolution layers are used as feature maps to reduce the number of model parameters. For remote sensing images in the front and back phases, after the feature extraction backbone network with shared parameters, multi-scale and multi-depth feature maps are generated respectively. , , , , , ;

[0024] (2) Feature Difference Module

[0025] In order to further obtain the change information, the difference between the two-phase feature maps generated by (1) is made, that is, These differential feature maps contain different scales and numbers of channels, and can provide ground cover change information of different values;

[0026] (3) Feature fusion module

[0027] The differential feature maps of different scales generated in (2) are bilinearly sampled into the input size and channel fused. Channel fusion can provide richer feature representation and finally generate the feature difference map F.

[0028] (4) Global attention module based on matrix decomposition

[0029] In order to solve the problem of high global attention overhead, the present invention inputs the feature map F after channel feature fusion (3) into a global information module, which is a matrix decomposition process, namely ; For convolutional neural networks, when an image is input, the network will output a tensor ; This tensor can be represented as a matrix , according to the matrix decomposition principle, is a linearly correlated low-rank matrix containing global information, and E is a noise matrix. The algorithm for solving the low-rank matrix is ​​encapsulated as a module structure of a neural network, which contains two linear transformations and a matrix decomposition model in the middle.

[0030] (5) Adaptive propagation module of pixel neighborhood

[0031] The feature map with global attention output in (4) The input is sent to this module, which can solve the problem of boundary loss after convolution. It is similar to a variable convolution kernel. It determines the pixel value of the center point by directly selecting the pixel values ​​of k points with variable distances in the surrounding area. In this way, each point gradually utilizes the pixels of the surrounding points and gradually expands to all points.

[0032] Furthermore, the model training, verification and testing evaluation methods are as follows:

[0033] 1) Model training;

[0034] 2) Model validation;

[0035] 3) Model testing and evaluation.

[0036] Furthermore, the model training:

[0037] The preprocessed data in step 2 is normalized and then input into the land change detection model in step 3 for training. The initial learning rate of the training is 0.01, and the learning rate is decayed according to the cosine function. The batch size is 8, the number of iterations is 100, the optimizer is SGD, and regularization technology is used to prevent overfitting of the network. At the same time, a weighted cross entropy function is used to measure the difference between the model prediction and the actual label, which can be expressed as:

[0038]

[0039] Where N is the number of samples, is the actual label, is the model’s predicted probability for the sample, are the weights of the changed class and the unchanged class, respectively.

[0040] Furthermore, the model verifies that:

[0041] Improve model performance by adjusting hyperparameters such as learning rate and batch size on the validation set.

[0042] Furthermore, the model was tested and evaluated:

[0043] The model of the present invention is evaluated using a test set. The evaluation indicators of the present invention include accuracy, precision, recall, F1, etc., and the formula is expressed as follows:

[0044] .

[0045] The present invention also provides a system for high-resolution remote sensing data processing and change detection, comprising:

[0046] Data preprocessing module for processing high-resolution remote sensing images, including image cropping and dataset segmentation;

[0047] The deep learning model building module is built on the Pytorch framework and adopts a twin differential structure to achieve end-to-end learning of remote sensing images in the previous and next phases.

[0048] Global attention module, based on matrix decomposition technology, is used to reduce the computational and memory overhead of global information modeling;

[0049] Adaptive propagation module, which solves the edge pixel sampling problem by learning a weighted average of local neighborhoods with fixed and variable two-dimensional offsets;

[0050] Training and verification module, used for model training, verification and performance evaluation;

[0051] The hardware environment includes an Intel(R) Core(TM) i9-10900X processor, an NVIDIA GeForce RTX 3080 graphics card, etc., running in Ubuntu 20.04 and Python 3.10 environment.

[0052] The present invention also provides a system for automated and accurate land use / land cover change detection, comprising a data acquisition module for multiple public remote sensing datasets, such as the WHU building change detection and LEVIR-CD datasets;

[0053] A preprocessing module that crops images into 256×256 slices and randomly allocates them to training / validation / test sets;

[0054] A deep learning model building module, combining a global attention module and an adaptive propagation algorithm, optimizes the model's ability to capture the edges of objects;

[0055] The training and validation module includes training, validation, and testing evaluation of the model on different datasets;

[0056] Computing and storage resource configuration modules, including high-performance computing processors and graphics cards, as well as an adapted development environment. Beneficial effects

[0057] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0058] First, the deep learning semantic segmentation model combined with matrix decomposition and adaptive propagation provided by the present invention is used to detect land changes. The model adopts a twin differential structure, which can simultaneously accept the input of pre-time and post-time remote sensing image data sets and share network parameters. Compared with traditional single-stream networks and post-classification-based methods, this structure has the advantage of end-to-end feature learning. Specifically, the model uses VGG-16 as the feature extraction network, and the dual-phase data generates feature maps of different scales after passing through the feature extraction network. In order to further obtain change information, pixel differences are made on the feature maps of different scales of the dual-phase, and then the differential feature maps of different scales are bilinearly upsampled to unify the shape and size, and channel fusion integrates the change feature information to capture change information of different scales and different depths.

[0059] This paper uses a matrix decomposition attention method to replace the traditional manual design attention method. In order to enable the model to better learn global features, a low-overhead global information module is introduced, which abstracts the global information into a mathematical model with a low-rank matrix. Assume that the input feature map can be represented as a matrix , F can be decomposed into the sum of a low-rank matrix and a noise matrix, that is ,in is a low-rank matrix with global information, and E is a noise matrix. Finally, the optimization algorithm for solving the low-rank matrix is ​​used as the structure of the neural network, which contains two linear transformations and a matrix decomposition model in the middle. Compared with other global attention methods, this attention module can avoid the manual design of weights or parameters. Secondly, the introduction of a low-overhead global information module effectively reduces the computational cost, making the model more suitable for large-scale data and high-resolution images. In addition, the compact representation of matrix decomposition helps to improve the robustness and adaptability of the model while suppressing the influence of noise in the global information. Most importantly, this method strikes a balance between computational efficiency and performance. By realizing the automatic learning of global information, it improves the model's ability to accurately capture real changes.

[0060] At the same time, in order to solve the problem of edge pixel sampling, the present invention combines the adaptive propagation technology of the adjacent pixel offset and introduces an adaptive propagation algorithm, that is, for each center pixel, it is calculated from a local neighborhood pixel. Specifically, the value of each pixel will be obtained by weighted average of the pixel values ​​of the neighborhood, but the neighborhood here is not a rectangular neighborhood in the conventional sense, but an offset is added on this basis, and this offset can be learned through the network. For the k domain points of each pixel point in the feature map, the module can learn the fixed two-dimensional offset D and the variable two-dimensional offset S of the k domain points, and calculate W through weighted interpolation, and the center pixel point The value of the neighborhood pixel P can be expressed as . This adaptive neighborhood selection can avoid sudden changes in pixel values ​​at the boundaries. Compared with other traditional edge detection algorithms, the adaptive propagation technology of the present invention has significant advantages. Traditional edge detection algorithms usually use fixed neighborhoods or manually designed rules for pixel sampling, which has limitations when dealing with complex edges and edges of objects. The adaptive propagation algorithm introduced in the present invention selects the neighborhood in a dynamic and adjustable manner by learning the fixed and variable offsets of the domain points, making it more flexible in dealing with various changes in the shape and structure of objects.

[0061] As described above, the present invention provides a deep learning semantic segmentation model that combines matrix decomposition and adaptive propagation, which can improve classification performance while ensuring the integrity of edge information, and realize automated and precise land use / land cover change detection on high-resolution optical data.

[0062] Second, the present invention belongs to the technical field of remote sensing land use / land cover change detection. To better achieve end-to-end change detection and simultaneously enable neural networks to learn global and edge information from feature maps at different scales and depths, the present invention proposes a deep learning semantic segmentation model that combines matrix decomposition and adaptive propagation, and applies it to optical remote sensing images. This model utilizes a twin differential architecture, which offers the advantages of end-to-end feature learning and can simultaneously accept input from both pre- and post-temporal remote sensing image datasets and share network parameters. Through twin dual-branch feature map differentials and channel fusion, rich multi-scale and multi-depth change information is learned. To further capture global information from feature maps, a matrix decomposition-based attention method is used instead of traditional manually designed attention methods. This method uses the low-rank matrix of the matrix decomposition as global information, and a neural network module is designed to solve the low-rank matrix. This method can learn global information from differential feature maps with low overhead and high performance. Furthermore, to address the edge pixel sampling issue, an adaptive propagation algorithm that incorporates neighboring pixel offsets is introduced. For each center pixel, a fixed offset and a variable offset are calculated from a local neighboring pixel. The adaptive propagation algorithm can solve the problem of boundary loss after convolution and avoid sudden changes in pixel values ​​at the boundary.

[0063] The land change detection method combining matrix decomposition and adaptive propagation proposed in the present invention can ensure the integrity of edge information while improving the performance of semantic classification, thereby improving the accuracy of land change detection, especially on high-resolution images, and providing more reliable data support for refined land use planning and environmental monitoring. Secondly, through the innovative introduction of matrix decomposition and adaptive propagation, the computational overhead of global attention is effectively reduced, the computational efficiency of the model is improved, and it is more suitable for processing large-scale high-resolution data. Furthermore, the introduction of the adaptive propagation module solves the problem of edge pixel sampling, improves the model's ability to capture the edges of objects, and makes land change detection more comprehensive and accurate.

[0064] Third, as auxiliary evidence for the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0065] (1) The expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0066] By more accurately identifying and quantifying land change, this invention enables decision-makers to obtain more scientific land resource utilization intelligence, helping to optimize urban planning, agricultural layout, and other areas, promoting sustainable development. Furthermore, customized land change detection services can be provided to meet the needs of different sectors. Government departments, environmental protection agencies, agricultural enterprises, and other organizations can obtain customized land use / cover change analysis reports based on their needs, providing more targeted data support for their decision-making.

[0067] (2) The technical solution of this invention fills the technical gap in the industry at home and abroad:

[0068] By optimizing feature learning, introducing a global attention mechanism, and employing an adaptive propagation algorithm, this method overcomes the over-focus on single pixels and excessive computational resource requirements of traditional land change detection methods, both domestically and internationally. This technical solution represents a significant breakthrough in remote sensing image processing, providing an efficient and accurate solution for land use / land cover change detection.

[0069] (3) The technical solution of the present invention overcomes technical prejudice:

[0070] This invention successfully overcomes the technical biases inherent in traditional land change detection methods, achieving more comprehensive, efficient, and accurate land change detection. Its unique twin differential architecture, matrix decomposition-based global attention module, and adaptive propagation module, among other technical innovations, bring new insights to the field of land change detection and provide more reliable and efficient technical support for related applications.

[0071] Fourth, the deep learning model of the present invention, which combines matrix decomposition and adaptive propagation, has achieved significant technical progress in high-resolution remote sensing data processing and change detection:

[0072] 1) Improved feature learning efficiency: By adopting a twin-differential architecture and sharing network parameters, the model significantly improves the efficiency of feature learning. This architecture enables the model to more accurately extract and compare features from remote sensing imagery in previous and subsequent time periods.

[0073] 2) Reduced computational and memory overhead: The introduced global attention module based on matrix decomposition effectively reduces the computational burden and memory consumption of global information modeling by abstracting global information into a low-rank matrix.

[0074] 3) Improved ability to capture object edges: By solving the edge pixel sampling problem through the adaptive propagation algorithm, the model can more accurately capture object edges, which is particularly important when processing high-resolution data.

[0075] 4) End-to-end learning capability: The model implements end-to-end learning from input remote sensing images to change detection output, simplifying the processing flow and improving processing efficiency.

[0076] 5) Improved model generalization: The model is trained and validated using multiple public datasets (such as the WHU building change detection and LEVIR-CD datasets) to improve generalization and practicality.

[0077] 6) Automated and accurate change detection: This method achieves automated and accurate land use / land cover change detection on high-resolution optical data, providing faster and more accurate detection results compared to traditional methods.

[0078] The technology of the present invention has made progress in feature learning efficiency, computational overhead, object edge capture capability, end-to-end learning, model generalization capability, and automated and accurate change detection, making it an effective tool for processing high-resolution remote sensing data. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] FIG1 is a flow chart of a land change detection method combining matrix decomposition and adaptive propagation provided by an embodiment of the present invention.

[0080] FIG2 is a flow chart of a land change detection method provided by an embodiment of the present invention.

[0081] FIG3 is a structural diagram of a model combining matrix decomposition and adaptive propagation provided by an embodiment of the present invention.

[0082] FIG4 is a global information module based on matrix decomposition provided by an embodiment of the present invention.

[0083] FIG5 is an adaptive propagation module for a pixel neighborhood provided by an embodiment of the present invention.

[0084] FIG6 is a diagram showing test results of the model of the present invention provided by an embodiment of the present invention.

[0085] FIG7 is a diagram showing the test results of the comparison models provided by an embodiment of the present invention on the LEVIR-CD dataset.

[0086] FIG8 is a diagram showing the test results of the comparative models provided in the embodiments of the present invention on the WHU dataset. Modes for Carrying Out the Invention

[0087] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0088] The land change detection method proposed in the present invention fully combines the techniques of matrix decomposition and adaptive propagation. By configuring a deep learning environment, acquiring and preprocessing high-resolution change detection data, and building a model of feature extraction, differentiation, fusion, global attention and adaptive propagation modules, accurate detection of land changes is achieved. During the model training, verification and testing process, the Pytorch deep learning framework was adopted, and two public datasets, WHU building change detection and LEVIR-CD, as well as specific hardware configurations and parameter settings, were used. After testing, the model of the present invention performed well on the LEVIR-CD and WHU building change detection datasets. In order to enable those skilled in the art to fully understand how the present invention is specifically implemented, this section is an explanatory embodiment that expands on the technical solution of the claims.

[0089] As shown in Figures 1 and 2, the present invention provides a land change detection method combining matrix decomposition and adaptive propagation, which includes the following steps:

[0090] S1, configure the deep learning environment;

[0091] This invention is based on the Pytorch open source deep learning framework, the operating environment is Ubuntu 20.04 version, the Python version is 3.10, the processor is Intel(R) Core(TM) i9-10900X, the graphics card is NVIDIA GeForce RTX 3080, and the development environment is PyCharm.

[0092] S2, high-resolution change detection data acquisition and preprocessing;

[0093] This paper uses two publicly available datasets: the WHU Building Change Detection (WHU) and the LEVIR-CD dataset. The WHU Building Change Detection dataset features a wide variety of buildings undergoing large-scale changes. The dataset consists of two cycles of aerial imagery and change labels, acquired in 2012 and 2016, respectively. Each image has a resolution of 0.3 meters. Since this dataset does not have a standard split, each image is cropped into 256×256 slices and randomly divided into training / validation / test sets in a ratio of 7:1:2, with no overlap between the slices. The LEVIR-CD dataset consists of 637 ultra-high-resolution image pairs collected from Google Earth and their corresponding change labels. Each image has a resolution of 0.5 meters and a size of 1024×1024. It is a large-scale change detection dataset covering a diverse range of buildings. The LEVIR-CD dataset provides a standard training / validation / test split, with 70% of the samples used for training, 10% for validation, and 20% for testing. Similarly, the samples are cropped into 256×256 slices.

[0094] S3, building a land change detection model;

[0095] (1) Feature extraction backbone network

[0096] As shown in Figure 3, the model of the present invention adopts a twin dual-branch structure with shared parameters, aiming to simultaneously learn the features of remote sensing images in the front and back phases. VGG-16 is used as the feature extraction backbone network, which consists of 16 convolutional layers and 3 fully connected layers. Multiple convolutional layers can learn multi-level feature representations. Since VGG-16 is large in model size and number of parameters, the results of the first three layers of convolution are used as feature maps to reduce the number of parameters of the model. For remote sensing images in the front and back phases, after the feature extraction backbone network with shared parameters, multi-scale and multi-depth feature maps are generated respectively. , , , , , .

[0097] (2) Feature Difference Module

[0098] In order to further obtain the change information, the difference between the two-phase feature maps generated by (1) is made, that is, These differential feature maps contain different scales and numbers of channels, and can provide different values ​​of ground cover change information.

[0099] (3) Feature fusion module

[0100] The differential feature maps of different scales generated by (2) are bilinearly sampled into the input size and channel fused. Channel fusion can provide richer feature representation and finally generate the feature difference map F.

[0101] (4) Global attention module based on matrix decomposition

[0102] In order to solve the problem of high global attention overhead, the present invention inputs the feature map F after channel feature fusion (3) into a global information module, which is a matrix decomposition process, namely For convolutional neural networks, when an image is input, the network will output a tensor . This tensor can be represented as a matrix , according to the matrix decomposition principle, is a linearly correlated low-rank matrix containing global information, and E is a noise matrix. As shown in Figure 4, the algorithm for solving the low-rank matrix is ​​encapsulated as a neural network module structure, which contains two linear transformations and a matrix decomposition model in the middle.

[0103] (5) Adaptive propagation module of pixel neighborhood

[0104] The feature map with global attention output in (4) The input is input into this module, as shown in Figure 5. This module can solve the problem of boundary loss after convolution. It is similar to a variable convolution kernel. It determines the pixel value of the center point by directly selecting the pixel values ​​of k points with variable distances in the surrounding area. In this way, each point gradually utilizes the pixels of the surrounding points and gradually expands to all points.

[0105] S4, model training, validation and test evaluation;

[0106] The present invention performs training, verification and testing in the deep learning environment described in step S1.

[0107] 1) Model training

[0108] The pre-processed data in step S2 is normalized and then input into the land change detection model in step S3 for training. The initial learning rate of the training is 0.01, the learning rate is decayed according to the cosine function, the batch size is 8, the number of iterations is 100, the optimizer is SGD, and regularization technology is used to prevent overfitting of the network. At the same time, a weighted cross entropy function is used to measure the difference between the model prediction and the actual label, which can be expressed as:

[0109]

[0110] Where N is the number of samples, is the actual label, is the model’s predicted probability for the sample, are the weights of the changed class and the unchanged class, respectively.

[0111] 2) Model Validation

[0112] Improve model performance by adjusting hyperparameters such as learning rate and batch size on the validation set.

[0113] 3) Model testing and evaluation

[0114] The model of the present invention is evaluated using the test set. The evaluation indicators of the present invention include accuracy, precision, recall, F1, etc., and the formula is expressed as follows:

[0115]

[0116]

[0117]

[0118]

[0119] After model testing, the proposed model achieved accuracy, precision, recall, and F1 of 98.90%, 87.60%, 90.57%, and 89.06% on the LEVIR-CD change detection dataset, and accuracy, precision, recall, and F1 of 99.26%, 90.39%, 94.38%, and 92.35% on the WHU building change detection dataset. The test results of the proposed model are shown in Figure 6. The proposed land change detection method, which combines matrix decomposition and adaptive propagation, improves semantic classification performance while ensuring the integrity of edge information.

[0120] The present invention provides a deep learning semantic segmentation model that combines matrix decomposition and adaptive propagation for detecting land changes. The model adopts a twin dual-branch structure, which can simultaneously accept the input of pre-time and post-time remote sensing image data sets and share network parameters. The structure has the advantage of end-to-end feature learning. Specifically, the model uses VGG-16 as the feature extraction network, and the dual-phase data generates feature maps of different scales after passing through the feature extraction network. In order to further obtain change information, pixel differences are made on the feature maps of different scales of the dual-phase, and then the differential feature maps of different scales are bilinearly upsampled to unify the shape and size, and channel fusion integrates the change feature information. The present invention uses a matrix decomposition attention method to replace the traditional manual design attention method, and combines the adaptive propagation technology of the adjacent point pixel offset to establish a deep learning twin differential semantic segmentation network to achieve automated and precise land use / land cover change detection on high-resolution optical data.

[0121] It should be noted that the embodiments of the present invention can be implemented through hardware, software, or a combination of software and hardware. The technical solutions of the present invention can be implemented on dedicated hardware, such as a specially designed remote sensing image processing chip. By optimizing the algorithm and model structure at the hardware level, the real-time performance and efficiency of land change detection can be improved. Hardware acceleration makes the method of the present invention more suitable for large-scale, high-resolution remote sensing data, meeting the needs of modern land management. The technical solutions of the present invention can be implemented through software, for example, by integrating the method into geographic information system (GIS) software. By embedding the method into existing GIS tools, users can easily utilize this advanced land change detection technology to achieve more accurate land use / land cover change analysis. The present invention can also be implemented through a combination of hardware and software. For example, by combining a high-performance computing platform with an optimized deep learning model, rapid and accurate land change detection can be achieved. In this embodiment, the hardware is responsible for accelerating the calculation, while the software is responsible for providing high-level algorithm and model control. Those skilled in the art will appreciate that the above-described devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such code being provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays and programmable logic devices, or by software executed by various types of processors, or by a combination of the above-described hardware circuits and software, such as firmware.

[0122] The present invention provides a deep learning semantic segmentation model that combines matrix decomposition and adaptive propagation for detecting land changes. The model adopts a twin dual-branch structure, which can simultaneously accept the input of pre-time and post-time remote sensing image data sets and share network parameters. The structure has the advantage of end-to-end feature learning. Specifically, the model uses VGG-16 as the feature extraction network, and the dual-phase data generates feature maps of different scales after passing through the feature extraction network. In order to further obtain change information, pixel differences are made on the feature maps of different scales of the dual-phase, and then the differential feature maps of different scales are bilinearly upsampled to unify the shape and size, and channel fusion integrates the change feature information. The present invention uses a matrix decomposition attention method to replace the traditional manual design attention method, and combines the adaptive propagation technology of the adjacent point pixel offset to establish a deep learning twin differential semantic segmentation network to achieve automated and precise land use / land cover change detection on high-resolution optical data.

[0123] The proposed model was compared with traditional and existing models and methods, such as FC-EF, FC-Siam-Di, FC-Siam-Conc, SNUNet, and IFNet. FC-EF is an image-level fusion method, in which bi-temporal images are concatenated into a single image and fed into a fully convolutional network. FC-Siam-Di is a feature-level fusion method, using a twin fully convolutional network to extract multi-level features and fusing bi-temporal information using simple feature differences. FC-Siam-Conc is also a feature-level fusion method, using a twin fully convolutional network to extract multi-level features and fusing bi-temporal information using feature concatenation. SNUNet is a multi-scale feature concatenation method that combines a twin network with UNet++ to extract high-resolution, high-level features. It uses channel-wise attention to concatenate features at each decoder level. IFNet is a multi-scale feature concatenation method that applies both channel-wise and spatial attention to concatenated temporal features at each decoder level. Comparative experiments on a high-resolution optical dataset demonstrated that the proposed deep learning semantic segmentation model performed well in the land change detection task. As shown in Tables 1 and 2, compared with other models, the classification accuracy of the present invention is significantly improved, and the model of the present invention can capture change information more accurately. As shown in Figures 7 and 8, the introduction of the adaptive propagation algorithm solves the problem of edge pixel sampling, effectively avoids the sudden change of pixel values ​​at the boundary, and improves the effect of semantic edge segmentation. Experimental results show that in land change detection in edge areas, the model of the present invention is more robust and accurate than the comparison model. At the same time, compared with the channel attention and spatial attention used in the comparison model, the global attention module of the present invention can automatically learn the global relationship in the image through methods such as matrix decomposition, without the need to manually design complex weights or rules, while channel attention and spatial attention require more manual design. At the same time, the global attention mechanism can establish connections across the entire input feature map, which helps the model to more comprehensively understand the context and global structure of the image. In contrast, channel attention focuses on modeling the relationship between channels, while spatial attention focuses on local areas. The introduction of global information enables the model to better capture the global patterns and associations in the image, thereby improving the perception of complex scenes and global structures.

[0124] Table 1 shows the test evaluation values ​​of the comparative models provided by the embodiments of the present invention on the LEVIR-CD dataset.

[0125]

[0126] Table 2 shows the test evaluation values ​​of the comparative models provided in the embodiments of the present invention on the WHU dataset.

[0127]

[0128] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A land change detection method combining matrix factorization and adaptive propagation, characterized in that, a deep learning model combining matrix factorization and adaptive propagation is used for land change detection of high-resolution remote sensing data; a siamese difference structure is used to achieve end-to-end learning of remote sensing images in different time phases, and the efficiency of feature learning is improved by sharing network parameters; a global attention module based on matrix factorization is used to reduce the computational and memory overhead of global information modeling, and at the same time, the edge pixel sampling problem is solved by an adaptive propagation algorithm to improve the model's ability to capture ground object edges; It is implemented based on the Pytorch framework, and the WHU building change detection and LEVIR-CD datasets are used to obtain, preprocess, build, train, validate and test high-resolution data.

2. The land change detection method combining matrix factorization and adaptive propagation according to claim 1, characterized in that, it specifically includes the following steps: Step 1, configure the deep learning environment; Step 2, obtain and preprocess high-resolution change detection data; Step 3, build a land change detection model; Step 4, model training, validation and test evaluation.

3. The land change detection method combining matrix factorization and adaptive propagation according to claim 2, characterized in that, the method for configuring the deep learning environment is as follows: Based on the open-source deep learning framework Pytorch, the running environment is Ubuntu20.04 version, the Python version is 3.10, the processor is Intel(R) Core(TM) i9-10900X, the graphics card is NVIDIA GeForce RTX 3080, and the development environment is PyCharm.

4. The land change detection method combining matrix factorization and adaptive propagation according to claim 2, characterized in that, the method for obtaining and preprocessing high-resolution change detection data: Two publicly available datasets, WHU building change detection and LEVIR-CD, are used; There are various large-scale changing buildings in the WHU building change detection dataset; This dataset consists of two-cycle aerial images and change labels, with the resolution of each image being 0.3 meters. Each image is cropped into image slices of 256×256 size, and they are randomly divided into training / validation / test sets in a ratio of 7:1:2, and there is no overlapping part in the slices during the allocation process; The LEVIR-CD dataset consists of 637 ultra-high-resolution image pairs collected from Google Earth and corresponding change labels, with the resolution of each image being 0.5 meters and the size being 1024×1024; The LEVIR-CD dataset provides a standard training / validation / test split, which uses 70% of the samples for training, 10% for validation, and 20% for testing; similarly, the samples are cropped into image slices of 256×256 size.

5. The land change detection method combining matrix factorization and adaptive propagation according to claim 2, characterized in that, the building of the land change detection model: (1) Feature extraction backbone network The model adopts a twin - branch structure with shared parameters and uses VGG - 16 as the feature extraction backbone network. This network consists of 16 convolutional layers and 3 fully - connected layers. Multiple convolutional layers can learn multi - level feature representations. The first three convolutional results are used as feature maps to reduce the number of model parameters. For the remote sensing images of the front and back double - temporal phases, after passing through the feature extraction backbone network with shared parameters, feature maps of multiple scales and multiple depths are generated respectively. , , , , , ； (2) Feature difference module Differentiate the feature maps of the front and back double time phases generated in (1), that is ； (3) Feature Fusion Module The differential feature maps of different scales generated in (2) are bilinearly sampled to the input size and channel fusion is performed. Channel fusion can provide a richer feature representation, and finally a feature difference map F is generated. (4) Global Attention Module Based on Matrix Decomposition The feature map F after fusing the features of (3) channels is input into a global information module; for a convolutional neural network, when an input image is provided, the network will output a tensor ; This tensor can be represented as a matrix , according to the matrix decomposition principle, is a low-rank matrix linearly related to global information, and E is a noise matrix; the algorithm for solving the low-rank matrix is encapsulated into the module structure of a neural network, which includes two linear transformations and a matrix decomposition model in the middle. (5) Adaptive Propagation Module for Pixel Neighborhood The feature map with global attention output in (4) Input into this module, the pixel value of the central point is determined by directly selecting the pixel values of k points with variable distances in the surrounding area. In this way, each point gradually utilizes the pixels of the surrounding points and gradually expands to all points.

6. The land change detection method combining matrix decomposition and adaptive propagation as described in claim 1, characterized in that the model training, validation and test evaluation method: 1) Model training; 2) Model validation; 3) Model testing and evaluation.

7. [Corrected according to Rule 26 on 14.12.2023] The land change detection method combining matrix decomposition and adaptive propagation as described in claim 5, characterized in that the model training: Normalize the preprocessed data in Step 2, and then input it into the land change detection model in Step 3 for training. The initial learning rate for training is 0.01, the learning rate decays according to the cosine function, the batch size is 8, the number of iterations is 100, the optimizer is SGD, and a regularization technique is adopted to prevent overfitting of the network; meanwhile, a weight-based cross-entropy function is used to measure the difference between the model prediction and the actual label, which can be expressed as: where N is the number of samples, and y i is the actual label, is the predicted probability of the model for the sample, w 1 and w 2 are the weights of the changed class and the unchanged class, respectively.

8. [Corrected according to Rule 26 on 14.12.2023] The land change detection method combining matrix decomposition and adaptive propagation as described in claim 5, characterized in that the model validation: By adjusting hyperparameters on the validation set to improve model performance; the model testing and evaluation: The model of the present invention is evaluated using a test set; the evaluation metrics of the present invention include accuracy, precision, recall, F1, etc., and the formula expressions are as follows:

9. A system for high-resolution remote sensing data processing and change detection, characterized in that it includes: A data preprocessing module for processing high-resolution remote sensing images, including image cropping and dataset segmentation; A deep learning model building module, constructed based on the Pytorch framework, adopting a siamese difference structure to achieve end-to-end learning of remote sensing images in different time phases; A global attention module, based on matrix decomposition technology, for reducing the computational and memory overhead of global information modeling; An adaptive propagation module for solving the problem of sampling edge pixel points, by learning the weighted average of local neighborhoods with fixed and variable two-dimensional offsets; A training and validation module for training, validating and evaluating the performance of the model.

10. A system for automated and precise land use / land cover change detection, characterized in that it includes a data acquisition module containing multiple publicly available remote sensing datasets, such as WHU building change detection and LEVIR-CD datasets; A preprocessing module for cropping the image into slices of 256×256 size and randomly allocating them to the training / validation / test sets; A deep learning model building module, combining a global attention module and an adaptive propagation algorithm to optimize the model's ability to capture ground object edges; A training and validation module, including training, validation and test evaluation of the model on different datasets; A computing and storage resource configuration module, including a high-performance computing processor and a graphics card, as well as a suitable development environment.

Images