Weakly supervised city change detection method and system fusing typical geographic element knowledge and generative training

Abstract

The invention relates to the technical field of remote sensing image change detection, in particular to a weak supervision city change detection method and system fusing typical geographic element knowledge and generative training, and the method comprises the steps: inputting a multi-temporal remote sensing image into a pre-trained change detection model, and obtaining a ground feature change region of a target region through the change detection model; the change detection model is constructed based on a weak supervision training framework of a generative adversarial network, and the generative adversarial network integrates a geographic element constraint rule under the weak supervision training framework to serve as prior knowledge of change detection to optimize a generator and a discriminator. The geographic element constraint rules comprise preset geographic element morphological constraint rules, phenological constraint rules and spatial constraint rules. According to the method, the change characteristics of real ground features can be more accurately captured, noise interference and seasonal influence are reduced, and the accuracy and reliability of a change detection result are improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image change detection technology, and in particular to a weakly supervised urban change detection method and system that integrates knowledge of typical geographic elements and generative training. Background Technology

[0002] Urban dynamic change detection is crucial for sustainable development. In the era of big data, massive amounts of remote sensing Earth observation data bring new opportunities and challenges to information extraction and identification. However, with rapid urban development, the demand for the timeliness of geographic information is increasing, making real-time change detection more urgent. Therefore, change detection of multi-temporal urban remote sensing imagery that takes into account geoscientific knowledge has always been a key focus and challenge in the field of surveying and mapping remote sensing.

[0003] In recent years, deep learning has played a significant role in remote sensing image interpretation, greatly improving automation and attracting increasing attention. However, it still has the following limitations: 1. Supervised learning relies on labeled data: Most current deep learning-based change detection methods are based on fully supervised learning, requiring training data to be labeled by professionals, consuming significant manpower and time, and resulting in high data processing costs. 2. Limited integration of remote sensing image change detection with geoscientific knowledge: Insufficient attention is paid to topography and spatial topological relationships between features. How to effectively utilize partially labeled data while considering the spatial constraints of real features to accurately capture image change features has become a problem that needs to be solved. Summary of the Invention

[0004] To address the spatiotemporal heterogeneity, noise interference, and seasonality issues inherent in multi-temporal remote sensing imagery, this invention provides a weakly supervised urban change detection method and system that integrates knowledge of typical geographic elements and generative training. It utilizes a small amount of labeled data and a large amount of unlabeled data to guide model training, reducing reliance on labeled data and lowering data acquisition costs. The morphological, phenological, and spatial constraint rules of four geographic elements—buildings, vegetation, water bodies, and roads—are embedded into the network training, overcoming the limitations of purely data-driven methods that often produce false changes that violate geographic laws. This maintains the spatiotemporal continuity of ground features and can be applied to fields such as urban planning, environmental monitoring, and disaster assessment.

[0005] According to the design scheme provided by this invention, on the one hand, a weakly supervised urban change detection method that integrates typical geographic element knowledge and generative training is provided, comprising:

[0006] Acquire multi-temporal remote sensing images of the target area;

[0007] Multi-temporal remote sensing images are input into a pre-trained change detection model, which is then used to obtain the areas of change in land features in the target region. The change detection model is constructed based on a weakly supervised training framework of generative adversarial networks (GANs). In this weakly supervised training framework, geographic element constraint rules are incorporated into the generator and discriminator as prior knowledge for change detection. The geographic element constraint rules include preset geographic element morphological constraint rules, phenological constraint rules, and spatial constraint rules.

[0008] As a weakly supervised urban change detection method that integrates typical geographic element knowledge and generative training in this invention, the geographic element morphological constraint rules further include: constraining the structure of land features in remote sensing images based on morphological operations to ensure the continuity and regularity of land feature changes. The morphological operations include: dilating land feature pixels and shrinking the dilated land feature pixels.

[0009] As a weakly supervised urban change detection method that integrates typical geographic element knowledge and generative training in this invention, the phenological constraint rules further include: using a multi-temporal co-constraint model to describe the coupling relationship between ground cover spectral characteristics and phenological changes, wherein the multi-temporal co-constraint model fits vegetation growth curves based on NDVI time series index and extracts key phenological parameters, and uses phenological phase difference as the basis for change detection.

[0010] As a weakly supervised urban change detection method that integrates typical geographic element knowledge and generative training according to the present invention, the spatial constraint rules further include: using three-level coupled decision-making and progressive filtering to fuse spatial rules and terrain features. The three-level coupled decision-making consists of spatial topology constraints, terrain adaptability, and spatial dynamic weight optimization. The spatial topology constraints are to perform topological consistency checks on remote sensing image features using a spatial relationship graph network. The terrain adaptability is to derive key terrain parameters based on the coupling constraint type rules of the features and terrain data and perform primary terrain filtering. The spatial dynamic weight optimization is to use factor weights to dynamically adjust the balance between spatial rule constraints and terrain feature preservation.

[0011] As a weakly supervised urban change detection method that integrates typical geographic element knowledge and generative training in this invention, the spatial constraint rule is further expressed as: F(x,y)=I topo (x,y)·I terrain (x,y)·P dynamic (x,y), where I topo (x,y) represents the binary indicator function of the spatial topological constraints of cell (x,y), I terrain (x,y) represents the terrain adaptation function for pixel (x,y), P dynamic (x,y) represents the dynamic weight adjustment factor of pixel (x,y).

[0012] As a weakly supervised urban change detection method that integrates typical geographic element knowledge and generative training, this invention further incorporates geographic element constraint rules as prior knowledge for change detection within the weakly supervised training framework, including:

[0013] Based on geographic element constraint rules, a hierarchical coding system is constructed with hard constraint layer and soft constraint layer. The hard constraint layer includes geographic element morphological constraint rules, and the soft constraint layer includes phenological constraint rules and spatial constraint rules.

[0014] Binary encoding is used to encode the constraint rules of the hard constraint layer, and the encoding result is used as the constraint term of the generator loss. Fuzzy logic is used to encode the constraint rules of the soft constraint layer, and the encoding result is used as the constraint term of the discriminator loss.

[0015] As a weakly supervised urban change detection method that integrates typical geographic element knowledge and generative training according to the present invention, the weakly supervised training process of the change detection model further includes:

[0016] Image data of known changed / pseudo-changed areas are selected from the remote sensing image dataset as training samples, and feature vectors are constructed based on the features of the image data;

[0017] Weak classifiers are trained using feature vectors to classify the entire remote sensing image in the remote sensing image dataset. Based on the classification results, pseudo-labels for the changes in land features in the corresponding remote sensing image data are generated.

[0018] In the pre-training stage, a rule encoder is trained using a remote sensing image dataset. The rule encoder performs hierarchical encoding of each geographic element constraint according to the importance and correlation of the constraint type, with hard constraint layer and soft constraint layer. A geoscience rule base composed of hard constraints and soft constraints is constructed. The hard constraints include morphological constraint rules of geographic elements in the target area, and the soft constraints include phenological constraint rules and spatial constraint rules of the target area. The rule encoder is described in the following text.

[0019] Hard constraint terms are added to the generator loss and soft constraint terms are added to the discriminator loss. During the training phase, the generator is guided to generate ground feature change detection results that conform to the geographical feature constraint rules through an adversarial game process between the generator and the discriminator, and the discriminator is prompted to accurately judge the authenticity of the ground feature change detection results.

[0020] Furthermore, this invention also provides a weakly supervised urban change detection system that integrates knowledge of typical geographical elements and generative training, comprising: a data acquisition module and a change detection module, wherein,

[0021] The data acquisition module is used to acquire multi-temporal remote sensing images of the target area;

[0022] The change detection module is used to input multi-temporal remote sensing images into a pre-trained change detection model, and use the change detection model to obtain the change areas of ground features in the target area. The change detection model is built based on a weakly supervised training framework of generative adversarial networks. In this weakly supervised training framework, the generative adversarial network incorporates geographic element constraint rules as prior knowledge for change detection to optimize the generator and discriminator. The geographic element constraint rules include preset geographic element morphological constraint rules, phenological constraint rules, and spatial constraint rules.

[0023] The beneficial effects of this invention are:

[0024] This invention integrates a knowledge base of typical geographic elements with a generative adversarial network (GAN), exemplified by GAN. By constructing morphological, phenological, and spatial constraint rules for four geographic elements—buildings, vegetation, water bodies, and roads—and employing a dual-path feature fusion strategy, these rules are embedded into the network training. This overcomes the limitations of purely data-driven methods, which often produce false changes that violate geographic laws, thus maintaining the spatiotemporal continuity of geographic features. Furthermore, a three-level coupled decision-making mechanism is used to construct a hierarchical coding system of "hard constraints-soft constraints," balancing rule constraints and detection accuracy. The model training is guided by a small amount of labeled data and a large amount of unlabeled data from remote sensing imagery at two different time points. The generative adversarial network enhances the model's generalization ability in complex scenarios, significantly reducing dependence on labeled data and improving the model's automation level and iteration efficiency in practical applications. Through experiments under multi-resolution, multi-seasonal, and noisy conditions, and by combining a spatiotemporal attention module to integrate spatial and temporal information, this solution strengthens feature focusing and background suppression in changed areas, effectively improving the model's robustness and detection accuracy. Attached image description:

[0025] Figure 1 This is a schematic diagram of a weakly supervised urban change detection process that integrates knowledge of typical geographic elements and generative training in the example.

[0026] Figure 2 This is a schematic diagram of the weakly supervised training process of the change detection model in the embodiment;

[0027] Figure 3 This is a schematic diagram of the generator network architecture in the embodiment;

[0028] Figure 4 This is a schematic diagram of the discriminator network architecture in the embodiment;

[0029] Figure 5 This is a schematic diagram illustrating the principle of integrating geographical knowledge with generative adversarial networks in the example.

[0030] Figure 6 This example illustrates the cross-scale generalization results of the dataset in the embodiment.

[0031] Figure 7 This is a schematic diagram of the model noise robustness verification results in the embodiment;

[0032] Figure 8 This is a schematic diagram of the anti-interference performance test results of the model in the embodiment;

[0033] Figure 9 This is a schematic diagram of the seasonal test results of the model in the example. Detailed implementation method:

[0034] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.

[0035] Current urban change detection methods typically rely on large amounts of pixel-level labeled data, resulting in high practical application costs and limited generalization ability. Furthermore, detection results based on spectral differences are often affected by sensor platforms and seasonal variations, leading to spurious changes and inaccurate results. Current technologies face the following key challenges:

[0036] (1) The problem of dependence on fully supervised learning: Most of the current mainstream change detection methods based on deep learning are based on fully supervised conditions. This means that the training data needs to be accurately labeled by professionals. This process not only consumes a lot of manpower and time, but also leads to high data processing costs. In practical applications, it is often very difficult to obtain large-scale accurately labeled training data, which limits the widespread application and promotion of these methods.

[0037] (2) Insufficient Integration of Geoscience Knowledge: Remote sensing image change detection is not adequately integrated with geoscience knowledge, and insufficient attention is paid to key information such as the spatial topological relationships between terrain and features. Multi-temporal remote sensing images suffer from spatiotemporal heterogeneity, noise interference, and seasonal effects. The lack of geoscience guidance makes it difficult for models to accurately capture the changing characteristics of real features, thus affecting the accuracy and reliability of change detection results. For example, in urban change detection, neglecting the spatial topological relationships between buildings and surrounding roads and green spaces may misjudge normal urban planning and construction changes as abnormal changes, or overlook some minor but important changes.

[0038] Therefore, in the embodiments of the present invention, see... Figure 1 As shown, a weakly supervised urban change detection method that integrates knowledge of typical geographic elements and generative training is presented, specifically including the following:

[0039] S101. Acquire multi-temporal remote sensing images of the target area;

[0040] S102. Input multi-temporal remote sensing images into a pre-trained change detection model, and use the change detection model to obtain the change areas of ground features in the target area. The change detection model is constructed based on a weakly supervised training framework of generative adversarial networks. In this weakly supervised training framework, the generative adversarial network incorporates geographic element constraint rules as prior knowledge for change detection to optimize the generator and discriminator. The geographic element constraint rules include preset geographic element morphological constraint rules, phenological constraint rules, and spatial constraint rules.

[0041] like Figure 2 As shown, model training is guided by a small amount of labeled data and a large amount of unlabeled data, reducing reliance on labeled data and data acquisition costs. A knowledge base integrating typical geographic elements and a generative adversarial network is used: morphological, phenological, and spatial constraint rules for four geographic elements—buildings, vegetation, water bodies, and roads—are embedded into the network training. This overcomes the limitations of purely data-driven methods, which often produce false changes that violate geographic laws, and maintains the spatiotemporal continuity of geographic features.

[0042] In this embodiment, the spatial variation patterns and constraints of typical urban features (such as buildings, roads, and green spaces) are analyzed in depth, and these rules are transformed into mathematical models. For example, changes in buildings usually follow certain planning principles; changes in their location, shape, and size must comply with urban planning requirements. The connection and direction of roads also exhibit certain regularities. By constructing spatial variation constraint models for these typical features, prior knowledge at the geographic element level is provided for change detection.

[0043] The morphological constraint rules for geographic elements include: constraining the structure of land features in remote sensing images based on morphological operations to ensure the continuity and regularity of land feature changes; the morphological operations include: dilating land feature pixels and shrinking the dilated land feature pixels.

[0044] By using structuring elements to expand the pixels of ground features in remote sensing images and then shrinking the expanded results, over-expanded pixels are eliminated, smooth ground feature outlines are preserved, and the smoothing effect is balanced between smoothing irregular shapes and avoiding misfilling. The smoothing effect can be quantified through evaluation metrics.

[0045] Taking Zhengzhou's urban area as an example, analysis and testing show that, regarding the continuity of building changes, new buildings must be adjacent to existing built-up areas and within 50 meters of existing buildings. In actual road reconstruction, more than 70% of the original road space is temporarily occupied or rebuilt in sections. Considering the evaluation indicators P and R, the boundary threshold θ is determined to be 0.7.

[0046] The phenological constraint rules include: using a multi-temporal synergistic constraint model to describe the coupling relationship between ground cover spectral characteristics and phenological changes; the multi-temporal synergistic constraint model fits vegetation growth curves based on the NDVI time series index and extracts key phenological parameters; and uses phenological phase difference as the basis for change detection.

[0047] The core of spectral-phenological constraints lies in the coupling relationship between the spectral characteristics of land cover and phenological changes. In this embodiment, a multi-temporal synergistic constraint model is constructed based on statistical analysis of multi-source remote sensing data and combined with ecological theory. Traditional remote sensing change detection methods based on NDVI time-series analysis only focus on amplitude changes, which can be affected by seasonal interference, leading to false detections. For example, the spectral characteristics of farmland during fallow periods are similar to those of bare soil from buildings, making it easy to misjudge as a change in land cover. In this embodiment, phenological phase difference (PPD) is introduced as the core criterion for change detection. PPD is defined as:

[0048]

[0049] in, This represents the phenological phase angle of pixel (x,y) at time t.

[0050] The NDVI time series index can be used to fit vegetation growth curves and extract key phenological parameters. Using an enhanced NDVI phenological constraint model incorporating PPD, the crop growing season is divided into four key stages: germination, dormancy, heading, and maturity. Key phenological periods are determined by the NDVI peak value, with each stage's phase corresponding to a specific angle. The core formula is:

[0051]

[0052] In the formula, the amplitude variation of NDVI maintains the original Gaussian distribution form, and the phase constraint uses the cosine square function to quantify the differences in phenological stages. When PPD = 0 (phenological synchronization), the weight is 1, and when PPD = π (completely out of phase), it drops to 0. If PPD > π / 4, that is, the phenological stage is ≥ 2 growth cycles, it is judged as an abnormal change, which can be further verified by combining spectral and morphological methods.

[0053] The spatial constraint rules include: using a three-level coupled decision and progressive filtering to fuse spatial rules and terrain features. The three-level coupled decision consists of spatial topology constraints, terrain adaptability, and spatial dynamic weight optimization. The spatial topology constraints are performed by using a spatial relationship graph network to check the topological consistency of remote sensing image features. The terrain adaptability is based on the coupled constraint type rules of the features and the terrain data to derive key terrain parameters and perform primary terrain filtering. The spatial dynamic weight optimization is to use factor weights to dynamically adjust the balance between spatial rule constraints and terrain feature preservation.

[0054] By utilizing a three-level coupled decision-making process and achieving effective integration of spatial rules and terrain features through progressive filtering, it can be expressed as: F(x,y)=I topo (x,y)·I terrain (x,y)·P dynamic (x,y), where I topo (x,y) represents the binary indicator function of the spatial topological constraints of cell (x,y), I terrain (x,y) represents the terrain adaptation function for pixel (x,y), P dynamic (x,y) represents the dynamic weight adjustment factor of pixel (x,y).

[0055] The three-level coupled decision-making process first performs primary terrain filtering, deriving key terrain parameters based on the coupling constraint type rules of land features and terrain data. Secondly, it constructs a spatial relationship graph network to perform topology consistency checks, and finally performs dynamic weight optimization.

[0056] Among them, incorporating geographic element constraint rules as prior knowledge for change detection within a weakly supervised training framework can be designed to include:

[0057] Based on geographic element constraint rules, a hierarchical coding system is constructed with hard constraint layer and soft constraint layer. The hard constraint layer includes geographic element morphological constraint rules, and the soft constraint layer includes phenological constraint rules and spatial constraint rules.

[0058] Binary encoding is used to encode the constraint rules of the hard constraint layer, and the encoding result is used as the constraint term of the generator loss. Fuzzy logic is used to encode the constraint rules of the soft constraint layer, and the encoding result is used as the constraint term of the discriminator loss.

[0059] Multiple constraints in the constructed typical land cover spatial change constraint model are hierarchically encoded. Different types of constraints are processed hierarchically according to their importance and relevance, so that these constraint information can be used more effectively in subsequent model training and change detection processes.

[0060] Based on change constraint rules for specific regions, a hierarchical rule coding strategy is adopted to construct a "hard constraint-soft constraint" hierarchical coding system, which takes into account the constraint rules while improving the quality of change detection. The hard constraint layer reflects the inviolable rules of natural laws and municipal planning, and is implemented using binary coding. Typical rules include topographic constraints and ecological protection zone constraints. The soft constraint layer is used to express the probabilistic rules of uncertainty and gradual change processes in geoscientific rules, and fuzzy logic is used to handle uncertainty.

[0061] Taking Generative Adversarial Networks (GANs) as an example, this paper optimizes the generator and discriminator in generative training. The generator is responsible for generating change detection results based on input multi-temporal remote sensing images and partial labeled data, while the discriminator is responsible for judging whether the generated results conform to the actual changes. By introducing hierarchical coding information from typical geographic feature spatial change constraint models, the generator and discriminator are trained and optimized, enabling the generator to generate change detection results that better conform to the rules of geographic feature change, and the discriminator to more accurately judge the authenticity of the results. A phased training strategy can be adopted during training, consisting of a pre-training phase and a joint fine-tuning phase. The pre-training phase uses a rule-only encoder for training, starting with flat area samples and gradually adding complex terrain to achieve end-to-end rule learning.

[0062] In the framework of generative adversarial training, the generative network learns the latent distribution of the input data to generate new data samples similar to real data. This allows it to compete with the discriminative network, prompting the network to continuously optimize its generation capabilities and produce higher-quality samples. The generative network architecture is as follows: Figure 3 As shown, the network consists of two branches. The input labeled data is divided into multiple land cover categories. The generator network generates corresponding image categories for each category of land cover in the image. For example, the generator subnetworks for categories c and d are denoted as Gc and Gd, respectively. The upper part of the network is the encoder, which takes a 512×512 image at time T1 and category labels as input. After a fourth convolutional layer and downsampling, the image size is reduced to 32×32, and the number of channels becomes 256. The lower part of the network is the decoder, which gradually restores the image size through upsampling operations. The final generated image T1' is 512×512 in size and has 3 channels.

[0063] The discriminative network is a core component of generative adversarial networks (GANs), forming the framework for adversarial training together with the generative network. Its core objective is to distinguish whether the input data comes from a real dataset or is generated adversarially, thereby continuously improving model performance. The discriminative network structure is as follows: Figure 4 As shown, T2 and T1' images with a size of 512×512 and 3 channels, along with category labels, are used as input. The discriminant network consists of three sub-networks sharing parameters. The discriminant sub-networks for categories c and d are denoted as Dc and Dd, respectively, and adopt the structure of a VGG network. The branches of the generator and transform networks provide greater flexibility for semantic segmentation and classification between images.

[0064] Specifically, the weakly supervised training process for the change detection model can be designed to include:

[0065] Image data of known changed / pseudo-changed areas are selected from the remote sensing image dataset as training samples, and feature vectors are constructed based on the features of the image data;

[0066] Weak classifiers are trained using feature vectors to classify the entire remote sensing image in the remote sensing image dataset. Based on the classification results, pseudo-labels for the changes in land features in the corresponding remote sensing image data are generated.

[0067] In the pre-training stage, a rule encoder is trained using a remote sensing image dataset. The rule encoder performs hierarchical encoding of each geographic element constraint according to the importance and correlation of the constraint type, with hard constraint layer and soft constraint layer. A geoscience rule base composed of hard constraints and soft constraints is constructed. The hard constraints include morphological constraint rules of geographic elements in the target area, and the soft constraints include phenological constraint rules and spatial constraint rules of the target area. The rule encoder is described in the following text.

[0068] Hard constraint terms are added to the generator loss and soft constraint terms are added to the discriminator loss. During the training phase, the generator is guided to generate ground feature change detection results that conform to the geographical feature constraint rules through an adversarial game process between the generator and the discriminator, and the discriminator is prompted to accurately judge the authenticity of the ground feature change detection results.

[0069] In this embodiment, the constructed geoscientific knowledge model is integrated with the optimized GAN. The geoscientific knowledge model provides prior knowledge about changes in urban landforms, while the GAN learns the change features in the data through generative training. This integration enables the model to perform change detection tasks under geoscientific constraints, fully utilizing geoscientific knowledge to guide the model's learning process and improving its adaptability to complex urban environments and spatiotemporal heterogeneity. Figure 5 As shown, a dual-path feature fusion strategy is used to embed the element rule base into the GAN training process in a multi-granular manner. First, low-level features are fused by injecting rule features into the convolutional layer of the discriminant network; second, high-level semantics are constrained by implementing rule consistency loss through an auxiliary classifier; finally, dynamic weight allocation is performed to adjust the importance of rules based on an attention mechanism.

[0070] Furthermore, based on the above method, this embodiment of the invention also provides a weakly supervised urban change detection system that integrates typical geographic element knowledge and generative training, comprising: a data acquisition module and a change detection module, wherein,

[0071] The data acquisition module is used to acquire multi-temporal remote sensing images of the target area;

[0072] The change detection module is used to input multi-temporal remote sensing images into a pre-trained change detection model, and use the change detection model to obtain the change areas of ground features in the target area. The change detection model is built based on a weakly supervised training framework of generative adversarial networks. In this weakly supervised training framework, the generative adversarial network incorporates geographic element constraint rules as prior knowledge for change detection to optimize the generator and discriminator. The geographic element constraint rules include preset geographic element morphological constraint rules, phenological constraint rules, and spatial constraint rules.

[0073] To verify the effectiveness of this solution, the following explanation is based on experimental data:

[0074] The experiment was conducted on a Windows 10 system using PyTorch framework version 3.7.14, with NVIDIA Quadro P5000 used for training. The change detection experiment used the Adam optimizer with an initial learning rate of 0.0002, dynamically adjusted during training, 200 iterations, and a batch size of 16. The UCM-GTK model was weakly supervised and trained according to these settings to improve model performance. Predictions were performed on the SECOND and WHU-GCD datasets under the same environment, and the detection metrics for different networks are shown in the table below.

[0075] Network comparison results for the SECOND dataset

[0076]

[0077]

[0078] Network comparison results for the WHU-GCD dataset

[0079]

[0080] By integrating multimodal geographic data for a unified representation, the model can comprehensively utilize multiple information sources, enhancing its adaptability to complex urban environments and spatiotemporal heterogeneity. Regardless of different geographical regions, seasonal conditions, or the presence of noise interference, the model maintains good change detection performance. The cross-scale generalization results of the two datasets are shown below. Figure 6 As shown. After adding common Gaussian noise (μ=0, σ=0.1), impulse noise (5% of pixels), and cloud / fog simulation to the dataset, the model's noise robustness verification results are as follows. Figure 7 As shown. The results of the anti-interference capability test are as follows. Figure 8 As shown. Seasonal test results are as follows. Figure 9As shown, the SDI of the SECOND dataset is slightly lower than that of WHU-GCD, but both reach above 0.91, proving that the proposed scheme GK-WSCD is more adaptable to seasonal changes.

[0081] The experimental data above demonstrates that the proposed solution employs a weakly supervised learning method, eliminating the need for precise labeling of all training data and requiring only a subset of labeled data for model training. This significantly reduces the workload and cost of data labeling, improving data processing efficiency. On public datasets, only 10% of the labeled samples are used to achieve an IoU of 65.43%, enabling effective urban change detection even with limited data availability. By constructing a typical spatial change constraint model for ground features and integrating geoscientific knowledge models with GANs, the spatial change patterns and constraints of urban ground features are fully considered. During change detection, the solution can more accurately capture the changing characteristics of real ground features, reducing noise interference and seasonal effects, thereby improving the accuracy and reliability of change detection results. This remote sensing intelligent interpretation method, combining geoscientific knowledge with deep learning generative training, not only provides an effective solution for urban change detection but also offers new ideas and references for remote sensing image interpretation in other fields.

[0082] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0083] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0084] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0085] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0086] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A weakly supervised urban change detection method that integrates knowledge of typical geographical elements and generative training, characterized in that, Includes: acquiring multi-temporal remote sensing images of the target area; Multi-temporal remote sensing images are input into a pre-trained change detection model, which is then used to obtain the areas of change in land features in the target region. The change detection model is constructed based on a weakly supervised training framework of generative adversarial networks (GANs). In this weakly supervised training framework, geographic element constraint rules are incorporated into the generator and discriminator as prior knowledge for change detection. The geographic element constraint rules include preset geographic element morphological constraint rules, phenological constraint rules, and spatial constraint rules.

2. The weakly supervised urban change detection method according to claim 1, which integrates knowledge of typical geographical elements and generative training, is characterized in that... The morphological constraint rules for geographic elements include: constraining the structure of land features in remote sensing images based on morphological operations to ensure the continuity and regularity of land feature changes. The morphological operations include: dilating land feature pixels and shrinking the dilated land feature pixels.

3. The weakly supervised urban change detection method according to claim 1, which integrates knowledge of typical geographical elements and generative training, is characterized in that... The phenological constraint rules include: using a multi-temporal synergistic constraint model to describe the coupling relationship between ground cover spectral characteristics and phenological changes. The multi-temporal synergistic constraint model fits vegetation growth curves based on the NDVI time series index and extracts key phenological parameters, and uses phenological phase difference as the basis for change detection.

4. The weakly supervised urban change detection method according to claim 1, which integrates knowledge of typical geographical elements and generative training, is characterized in that... The spatial constraint rules include: using a three-level coupled decision and progressive filtering to fuse spatial rules and terrain features. The three-level coupled decision consists of spatial topology constraints, terrain adaptability, and spatial dynamic weight optimization. The spatial topology constraints are performed by using a spatial relationship graph network to check the topological consistency of remote sensing image features. The terrain adaptability is based on the coupled constraint type rules of the features and the terrain data to derive key terrain parameters and perform primary terrain filtering. The spatial dynamic weight optimization is to use factor weights to dynamically adjust the balance between spatial rule constraints and terrain feature preservation.

5. The weakly supervised urban change detection method according to claim 4, which integrates knowledge of typical geographical elements and generative training, is characterized in that... The spatial constraint rule is expressed as: F(x,y)=I topo (x,y)·I terrain (x,y)·P dynamic (x,y), where I topo (x,y) represents the binary indicator function of the spatial topological constraints of cell (x,y), I terrain (x,y) represents the terrain adaptation function for pixel (x,y), P dynamic (x,y) represents the dynamic weight adjustment factor of pixel (x,y).

6. The weakly supervised urban change detection method according to claim 1, which integrates knowledge of typical geographical elements and generative training, is characterized in that... Integrating geographic feature constraint rules as prior knowledge for change detection within a weakly supervised training framework, including: Based on geographic element constraint rules, a hierarchical coding system is constructed with hard constraint layer and soft constraint layer. The hard constraint layer includes geographic element morphological constraint rules, and the soft constraint layer includes phenological constraint rules and spatial constraint rules. Binary encoding is used to encode the constraint rules of the hard constraint layer, and the encoding result is used as the constraint term of the generator loss. Fuzzy logic is used to encode the constraint rules of the soft constraint layer, and the encoding result is used as the constraint term of the discriminator loss.

7. The weakly supervised urban change detection method according to claim 1, which integrates knowledge of typical geographical elements and generative training, is characterized in that... The weakly supervised training process of the change detection model includes: Image data of known changed / pseudo-changed areas are selected from the remote sensing image dataset as training samples, and feature vectors are constructed based on the features of the image data; Weak classifiers are trained using feature vectors to classify the entire remote sensing image in the remote sensing image dataset. Based on the classification results, pseudo-labels for the changes in land features in the corresponding remote sensing image data are generated. In the pre-training stage, a rule encoder is trained using a remote sensing image dataset. The rule encoder performs hierarchical encoding of each geographic element constraint according to the importance and correlation of the constraint type, with hard constraint layer and soft constraint layer. A geoscience rule base composed of hard constraints and soft constraints is constructed. The hard constraints include morphological constraint rules of geographic elements in the target area, and the soft constraints include phenological constraint rules and spatial constraint rules of the target area. The rule encoder is described in the following text. Hard constraint terms are added to the generator loss and soft constraint terms are added to the discriminator loss. During the training phase, the generator is guided to generate ground feature change detection results that conform to the geographical feature constraint rules through an adversarial game process between the generator and the discriminator, and the discriminator is prompted to accurately judge the authenticity of the ground feature change detection results.

8. A weakly supervised urban change detection system that integrates knowledge of typical geographical elements and generative training, characterized in that, It includes: a data acquisition module and a change detection module, wherein, The data acquisition module is used to acquire multi-temporal remote sensing images of the target area; The change detection module is used to input multi-temporal remote sensing images into a pre-trained change detection model, and use the change detection model to obtain the change areas of ground features in the target area. The change detection model is built based on a weakly supervised training framework of generative adversarial networks. In this weakly supervised training framework, the generative adversarial network incorporates geographic element constraint rules as prior knowledge for change detection to optimize the generator and discriminator. The geographic element constraint rules include preset geographic element morphological constraint rules, phenological constraint rules, and spatial constraint rules.

9. An electronic device, characterized in that, include: At least one processor, and a memory coupled to said at least one processor; The memory stores a computer program that can be executed by the at least one processor to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, enables the implementation of the method as described in any one of claims 1 to 7.

Images