Satellite image ground object change detection method and system based on long time sequence information fusion

Abstract

The invention discloses a satellite image ground object change detection method and system based on long time sequence information fusion, and relates to the technical field of remote sensing image processing. A data acquisition module collects remote sensing image multi-source data of the same geographic area at different times; the data preprocessing module carries out preprocessing; a multi-modal feature extraction module extracts high-level semantic features from the preprocessed remote sensing image multi-source data; the time sequence information fusion module carries out deep space-time fusion on the high-level semantic features extracted at each time point in the time dimension to obtain deep space-time fusion features; a change judgment module performs ground feature category change analysis according to the deep space-time fusion features, marks the area where the ground feature category changes, and generates a change detection result of the geographic area; and the result output and feedback module outputs the change detection result to a user terminal for display and storage, and receives verification and correction information of the user on the change detection result for model training updating.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method and system for detecting ground feature changes in satellite images based on long-term time-series information fusion. Background Technology

[0002] With the rapid development of remote sensing technology, the ability to observe the Earth has been greatly enhanced. Remote sensing imagery has seen significant improvements in spatial, temporal, and spectral resolution, leading to its widespread application in urban planning, environmental monitoring, and disaster assessment. Particularly in land cover change detection, acquiring dynamic geographic information using intelligent remote sensing technology has become a crucial method. Change detection, as one of the core tasks in remote sensing applications, aims to identify changes in land cover types over time by analyzing remote sensing images from different temporal phases, providing scientific basis and decision support for land use management and ecological environmental protection.

[0003] Traditional remote sensing semantic change detection methods typically employ a "two-temporal classification and comparison" strategy: pixel-by-pixel semantic classification is performed on two consecutive remote sensing images, assigning each pixel to a specific land cover category, such as water bodies, vegetation, or built-up areas. After completing the classification for each temporal phase, the classification results from the two phases are compared pixel-by-pixel to identify areas of land cover change and the type of change. This method heavily relies on the classification accuracy of each image: if a particular temporal image has a significant classification error due to cloud cover, lighting variations, or sensor noise, it may lead to false positives or false negatives of changed areas, affecting the reliability of the final detection results. Furthermore, traditional methods often only focus on the differences between two images, ignoring the inherent correlations between multiple temporal images, making it difficult to effectively capture subtle, gradual changes in the land surface. For slow-moving changes such as gradual vegetation degradation or urban expansion, which require long-term monitoring, the "two-temporal" method is often inadequate. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting ground feature changes in satellite images based on long-term time-series information fusion, which solves the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Satellite imagery-based methods for detecting ground feature changes based on long-term time-series information fusion include: Collect multi-source remote sensing imagery data of the same geographic area at different times; Preprocessing is performed on the multi-source remote sensing imagery acquired at different times; Based on a deep learning model, high-level semantic features are extracted from preprocessed remote sensing image multi-source data. The high-level semantic features extracted at each time point are deeply spatiotemporally fused along the time dimension to obtain deep spatiotemporally fused features, including: The correlation weights between high-level semantic features at different times are calculated at each spatial location using a temporal attention mechanism, and the fused features are obtained through attention weighting. The high-level semantic features of each frame of the time series image are introduced into the graph neural network as nodes in the time graph sequence to model the relationship of the temporal features. The graph neural network node represents the high-level semantic features of a certain image block or pixel position at a specific time. Temporally adjacent nodes and spatially adjacent regions are connected by edges. Through iterative information transmission, the graph neural network enables the high-level semantic features of adjacent times and adjacent spatial positions to influence each other, achieving deep fusion of spatiotemporal features and obtaining deep spatiotemporal fusion features. Based on the deep spatiotemporal fusion characteristics, we perform land cover category change analysis, mark the areas where land cover category changes have occurred through change analysis, and generate change detection results for geographic areas. The system outputs the change detection results of the geographical area to the user terminal for display and storage, receives the user's verification and correction information on the change detection results, and uses the verification and correction information for model training and updates.

[0006] Furthermore, the preprocessing of the multi-source remote sensing imagery acquired at different times includes: Aligning multi-source remote sensing image data from different times to the same coordinate system enables geometric correction and registration of multi-source remote sensing image data; For optical images, a radiometric correction algorithm based on an atmospheric radiative transfer model is used to eliminate the effects of atmospheric scattering and absorption, remove cloud and fog masking, and achieve radiometric correction. Perform multi-view processing and filtering / denoising operations on SAR images; The image resolution is adjusted to a fixed value using bilinear interpolation or nearest neighbor interpolation. For multimodal data, registration and resolution matching are performed between different data sources, and a unified format conversion is carried out.

[0007] Furthermore, the specific process for extracting high-level semantic features is as follows: For single-modal images, a convolutional neural network combined with a visual Transformer encoder is used to extract spatial features, including: Convolutional neural networks with a predetermined number of convolutional stages are constructed based on ResNet-50; After each convolutional stage of the convolutional neural network, the corresponding feature map is obtained, and a feature pyramid network is used to fuse multi-scale features. The high-level features are upsampled and added to the low-level features element by element. The output feature map of the last convolutional stage after the feature pyramid network is fused is input into the visual Transformer encoder to output high-level semantic features of a single modality. For multimodal data, a multi-branch feature extraction model or a unified fusion network feature extraction model is used to extract features from different modalities of the image, including: Each modality is set up with an independent branch. After each branch extracts features according to the corresponding architecture, a contrastive learning strategy is introduced to calculate the cosine similarity of features of different modalities at the same position. A contrastive loss function is constructed. When the similarity is lower than the set similarity threshold, the network parameters of each branch are updated through backpropagation to make the feature representations of different modalities at the same position more semantically consistent. A hierarchical fusion strategy is adopted to perform modality fusion in the early, middle and late stages of the fusion network. The fused features are then used to extract high-level semantic features through subsequent convolutional layers, pooling layers and fully connected layers.

[0008] Furthermore, the deep fusion achieving spatiotemporal features includes: For the extracted high-level semantic features at each time point, a Transformer-based multi-head self-attention structure is applied at each spatial location to calculate the correlation weights between high-level semantic features at different times. The high-level semantic features of each frame of the time series are regarded as nodes in the time graph sequence. Temporally adjacent nodes and spatially adjacent regions are connected by edges to construct a spatiotemporal graph. Through iterative information transmission, high-level semantic features of adjacent time points and adjacent spatial locations influence each other, achieving deep fusion of spatiotemporal features; The fused spatiotemporal features are subjected to dimensionality reduction and normalization.

[0009] Furthermore, the step of performing land cover category change analysis based on deep spatiotemporal fusion features, marking areas where land cover category changes have occurred through change analysis, and generating change detection results for geographic areas includes: The decoder employs layer-by-layer upsampling to restore the temporally fused features to the original image resolution. During the decoding process, cross-layer skip connections are used to introduce high-resolution spatial details, and fine semantic segmentation results are output for each time frame. The system performs change analysis on the selected baseline time phase and the current time phase. By comparing the semantic segmentation maps at different times, or by directly outputting the change mask from the model, it determines which areas have experienced changes in land cover categories. It identifies the location range of the changed areas, combines semantic information to provide a category description of the change as a label, and outputs the labeled change results.

[0010] This invention also provides a satellite image land cover change detection system based on long-term time-series information fusion, comprising: The data acquisition module is used to collect multi-source remote sensing imagery data of the same geographical area at different times; The data preprocessing module is used to preprocess multi-source remote sensing image data acquired at different times; The multimodal feature extraction module is used to extract high-level semantic features from preprocessed remote sensing image multi-source data based on a deep learning model. The temporal information fusion module is responsible for performing deep spatiotemporal fusion of the high-level semantic features extracted from each time point in the time dimension to obtain deep spatiotemporal fused features, including: The correlation weights between high-level semantic features at different times are calculated at each spatial location using a temporal attention mechanism, and the fused features are obtained through attention weighting. The high-level semantic features of each frame of the time series image are introduced into the graph neural network as nodes in the time graph sequence to model the relationship of the temporal features. The graph neural network node represents the high-level semantic features of a certain image block or pixel position at a specific time. Temporally adjacent nodes and spatially adjacent regions are connected by edges. Through iterative information transmission, the graph neural network enables the high-level semantic features of adjacent times and adjacent spatial positions to influence each other, achieving deep fusion of spatiotemporal features and obtaining deep spatiotemporal fusion features. The change determination module is used to perform change analysis of land cover categories based on deep spatiotemporal fusion features. Through change analysis, it marks areas where land cover categories have changed and generates change detection results for geographic areas. The results output and feedback module is used to output the change detection results of the geographical area to the user terminal for display and storage, receive the user's verification and correction information on the change detection results, and use the verification and correction information for model training and updates.

[0011] Furthermore, the data preprocessing module includes: The geometric correction submodule is used to align multi-source remote sensing image data from different times to the same coordinate system, thereby achieving geometric correction and registration of multi-source remote sensing image data. The radiometric correction submodule is used to eliminate the effects of atmospheric scattering and absorption, remove cloud and fog masks, and achieve radiometric correction for optical images using a radiometric correction algorithm based on the atmospheric radiative transfer model. The denoising submodule is used for multi-view processing and filtering / denoising operations on SAR images; The unified resolution submodule is used to adjust the image resolution to a set fixed value using bilinear interpolation or nearest neighbor interpolation. The format conversion submodule is used to perform registration and resolution matching between different data sources for multimodal data, and to perform unified format conversion.

[0012] Furthermore, the multimodal feature extraction module includes: The single-modality image feature extraction submodule is used to extract spatial features from single-modality images using a convolutional neural network combined with a visual Transformer encoder, including: Convolutional neural networks with a predetermined number of convolutional stages are constructed based on ResNet-50; After each convolutional stage of the convolutional neural network, the corresponding feature map is obtained, and a feature pyramid network is used to fuse multi-scale features. The high-level features are upsampled and added to the low-level features element by element. The output feature map of the last convolutional stage after the feature pyramid network is fused is input into the visual Transformer encoder to output high-level semantic features of a single modality. The multimodal feature extraction submodule is used to extract features from different modalities of images using a multi-branch feature extraction model or a unified fusion network feature extraction model, including: Each modality is set up with an independent branch. After each branch extracts features according to the corresponding architecture, a contrastive learning strategy is introduced to calculate the cosine similarity of features of different modalities at the same position. A contrastive loss function is constructed. When the similarity is lower than the set similarity threshold, the network parameters of each branch are updated through backpropagation to make the feature representations of different modalities at the same position more semantically consistent. A hierarchical fusion strategy is adopted to perform modality fusion in the early, middle and late stages of the fusion network. The fused features are then used to extract high-level semantic features through subsequent convolutional layers, pooling layers and fully connected layers.

[0013] Furthermore, the time-series information fusion module includes: The temporal attention mechanism processing submodule is used to apply a Transformer-based multi-head self-attention structure to calculate the correlation weights between high-level semantic features at different times for the extracted high-level semantic features at each spatial location. The spatiotemporal graph construction submodule is used to treat the high-level semantic features of each frame of the time series as nodes in the time graph sequence. Temporally adjacent nodes and spatially adjacent regions are connected by edges to construct the spatiotemporal graph. The spatiotemporal fusion submodule is used to achieve deep fusion of spatiotemporal features by enabling high-level semantic features of adjacent time points and adjacent spatial locations to influence each other through iterative information transmission. The feature post-processing submodule is used to perform dimensionality reduction and normalization on the fused spatiotemporal features.

[0014] Furthermore, the change determination module includes: The decoding submodule is used to use the decoder to upsample layer by layer to restore the temporally fused features to the original image resolution. During the decoding process, cross-layer skip connections are used to introduce high-resolution spatial details and output fine semantic segmentation results for each time frame. The decision submodule is used to perform change analysis on the selected reference time phase and the current time phase. By comparing the semantic segmentation maps at different times, or by directly outputting the change mask from the model, it determines which areas have changed the land cover categories. It identifies the location range of the changed areas, combines semantic information to give a category description of the change as a label, and outputs the labeled change results.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention overcomes the limitations of traditional methods that rely solely on comparing two images. It fully utilizes the characteristics of the same land cover's evolution over time within multiple remote sensing images, improving the accuracy and robustness of change detection by introducing a long-term time-series information fusion mechanism. Specifically, this invention performs joint analysis on remote sensing images of the same region acquired at multiple different times: First, it identifies and classifies land cover elements in the images at each time point, obtaining initial semantic segmentation results for each temporal phase. Then, instead of simply comparing the classification images from two different times, it designs a deep learning-driven time-series feature fusion and change discrimination model. This model comprehensively considers the spectral characteristics, semantic classification results, and evolution trends of a given geographical location at multiple historical times to determine whether a substantial change in land cover type has occurred at that location, and the nature or type of the change.

[0016] The method of this invention integrates several innovative technologies in its algorithm, including but not limited to: temporal attention mechanism, graph neural network modeling, cross-modal contrastive learning, and self-supervised feature extraction. The temporal attention mechanism enables the model to automatically focus on regions that have truly changed over a long time axis, ignoring short-term noise interference; the introduction of graph neural networks (GNNs) allows the model to construct spatiotemporal relationship graphs using pixels / regions as nodes, propagating information spatially and temporally, thereby capturing the correlation changes between adjacent regions and improving the coherence and accuracy of change detection; cross-modal contrastive learning enables the system to fuse multi-source remote sensing data (e.g., optical imagery and SAR radar imagery), improving the model's feature representation ability across different data sources by performing feature alignment and consistency constraints on different modalities, reducing the limitations of a single sensor; self-supervised feature extraction utilizes massive amounts of unlabeled remote sensing data to pre-train the model, learning general representations under unsupervised signals, and then transferring these representations for semantic segmentation and change detection tasks, enhancing feature robustness and reducing reliance on extensive manual annotation.

[0017] This invention efficiently acquires and processes long-term, multi-temporal remote sensing image data, automatically extracts discriminative spatiotemporal features, and intelligently identifies the areas and types of changes in land cover. Compared with traditional methods, this approach fully utilizes the redundant information of time series data, effectively reducing the impact of poor single-period image quality on the results. Furthermore, through innovative algorithm module combinations, it improves the sensitivity and detection accuracy for slow and subtle changes. This system is particularly suitable for scenarios requiring long-term continuous monitoring (such as urban expansion, forest succession, and changes in agricultural land use), enabling high-precision, automated change detection and providing new technical support for related fields. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the satellite image land cover change detection method based on long-term time-series information fusion according to the present invention; Figure 2 This is a schematic diagram of the satellite image ground feature change detection system based on long-term time-series information fusion according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To resolve existing technical issues, please refer to Figure 1 This embodiment provides the following technical solution: Satellite imagery-based methods for detecting ground feature changes based on long-term time-series information fusion include: Raw image acquisition: Collect multi-source remote sensing imagery of the same geographic area at different times from remote sensing data sources; for example, acquire multi-period optical satellite images of the area over many years, supplemented by other modal data such as radar images when necessary; the time interval for acquisition can be determined according to monitoring needs (such as monthly, quarterly or annually); it is necessary to ensure that the image coverage area is consistent or can be registered to be consistent, and select cloudless and high-resolution images as much as possible to improve the quality of analysis; Data preprocessing: Preprocessing is performed on multi-source remote sensing imagery acquired at different times to eliminate the influence of invariant factors on image differences. First, geometric correction and image registration are performed, resampling all images to the same map projection and coordinate system, and aligning them to the sub-pixel level to ensure that the same ground feature corresponds to the same location in images from different time periods. Next, radiometric correction / normalization is performed to balance brightness and color differences between images (e.g., eliminating overall brightness variations caused by different solar altitude angles and sensor gains), ensuring that brightness differences reflect real-world changes rather than changes in imaging conditions. For optical images, cloud and shadow artifacts should be identified and removed: cloud detection algorithms can be used to generate cloud masks, removing clouded areas or supplementing them with data from other clear time periods. In multi-source data scenarios, data from different modalities are preprocessed separately and resampled to the same spatial resolution and grid for subsequent fusion analysis. After this step, a set of high-quality, aligned time-series remote sensing images is output, providing reliable input for feature extraction. Feature Extraction and Representation Learning: Based on a deep learning model, high-level semantic features are extracted from preprocessed remote sensing imagery from multiple sources. In this embodiment, a model combining a trained convolutional neural network and a Transformer is used as the feature extractor (e.g., an encoder-decoder architecture, where the encoder extracts features and the decoder outputs preliminary classification results). The model first independently extracts multi-level feature maps for each image: low-level texture features, mid-level shape features, and high-level semantic features are obtained through layer-by-layer convolution and downsampling. To enhance the utilization of information in the temporal dimension, time position encoding is introduced during the encoding process, enabling the model to recognize the temporal order of the features. For each image, the model can output a... The initial semantic segmentation results (the probability distribution of each pixel belonging to a certain class) and a series of intermediate feature representations are used. The core purpose of this step is to transform the original image into a high-level representation that is easy to compare and fuse. It is worth emphasizing that self-supervised feature learning plays a role here: the model has been pre-trained with a large amount of unlabeled image data before it is formally executed. For example, by designing pre-training tasks such as predicting future image frames, temporal sorting, and cross-modal matching, the model has learned basic spatiotemporal patterns and cross-modal correspondences. After self-supervised pre-training and fine-tuning on a small number of labeled samples, the model can extract robust features, effectively alleviate the appearance differences caused by image noise and changes in lighting conditions, and lay a good foundation for subsequent change detection. Spatiotemporal Feature Fusion and Change Recognition: High-level semantic features extracted from each time point are deeply spatiotemporally fused along the time dimension to obtain deep spatiotemporal fused features. Specifically, a temporal attention mechanism (a multi-head self-attention structure based on Transformer) is used to calculate the correlation weights between high-level semantic features at different times at each spatial location, allowing the model to automatically "pay attention" to regions that change significantly along the time axis. Unlike traditional methods of compressing temporal information, this module retains an independent semantic representation for each time point and obtains fused features through attention weighting while maintaining complete temporal dimension information. High-level semantic features of each frame of the time series image are introduced into the graph neural network as nodes in the time graph sequence. The system models the relationships between temporal features. Graph neural network nodes represent high-level semantic features of an image patch or pixel location at a specific time. Temporally adjacent nodes and spatially adjacent regions are connected by edges. Through iterative information transmission, the graph neural network allows high-level semantic features of adjacent times and spatial locations to influence each other, achieving deep fusion of spatiotemporal features and obtaining deep spatiotemporal fusion features. An improved time attention mechanism combined with graph neural networks is used to model long-term series data. This fusion method can capture long-term dependencies and spatial neighborhood consistency, ensuring that changes in a city area in adjacent years are considered as a related whole, rather than scattered point changes. Through the synergistic effect of time attention and graph neural networks, this module outputs fully fused deep features, providing a reliable basis for change judgment. In this process, the system comprehensively applies time attention mechanisms and graph neural networks to mine change signals. Under the action of the time attention mechanism, the model calculates the correlation of feature sequences at each spatial location at different times: features at times that show abnormalities or significant changes in the time series are given higher weights, while features at stable and unchanged times are given relatively lower weights. This attention-weighted approach allows the model to focus on the truly changing time periods and regions, mitigating the interference of short-term anomalies (such as classification anomalies caused by cloud shadows at a certain time). For example, if a piece of land has been forested for most years but has recently become bare land, the attention mechanism will focus on the abrupt change in recent features at that location, thus determining that the area is likely to have experienced deforestation. Next, the fusion module constructs a spatiotemporal graph: treating the features of each pixel (or superpixel / region) at each time as graph nodes, connecting nodes of the same pixel at different times in the temporal dimension, and connecting nodes of adjacent pixels / regions at the same time in the spatial dimension, forming a spatiotemporal graph structure; through iterative computation of the graph neural network, nodes transmit and share information—for example, if adjacent pixels have experienced similar spectral changes at a certain time, the graph network will amplify the signal of this coherent change; or, nodes at adjacent times before and after the same location will mutually verify the persistence of the change.This graph structure learning can capture the changing spatial connectivity and temporal persistence, making the detection results more consistent and reliable. After multiple rounds of temporal attention fusion and graph network propagation, the model generates a deep feature representation that integrates the information of the entire time series, in which the pattern features that "change" and the background features that "remain unchanged" have been explicitly or implicitly distinguished. Change Information Extraction and Output: Based on deep spatiotemporal fusion features, land cover category change analysis is performed. Areas where land cover category changes have occurred are marked through this analysis, and change detection results for geographic areas are generated based on the fused spatiotemporal features. Specifically, firstly, the fused features are reconstructed using a decoder, upsampled to restore spatial resolution, and then fine-grained semantic segmentation results for each temporal phase are output. This means that while fully utilizing information from other temporal phases, the model still provides an independent land cover category map for each moment. These results are more accurate and consistent than classifying each period's image separately, providing a high-quality foundation for subsequent change comparison. Next, the system performs a change comparison analysis on two selected time periods (usually using the initial monitoring period as the base period, compared with the current / latest time phase; cumulative changes over the entire time period can also be detected if necessary). By comparing the base period and the current semantic segmentation map pixel by pixel, the set of pixels where the category has changed is marked, generating change patches. Each change patch is assigned attribute information, including its category in the base period and the current time phase, such as "changed from category X to category Y". Simultaneously, for gradual and slow changes (such as a year-on-year decrease in vegetation cover), the system can identify the overall trend of change (such as a pattern of "continuous vegetation degradation") by analyzing results from multiple periods and classify it into the corresponding change type. Finally, the change detection results are output in an easy-to-understand format, commonly a change map of the same size as the original image, where unchanged areas are marked "unchanged", and changed areas are colored according to their new semantic category or change type. For example, in a change map output from a city expansion scenario, newly constructed urban building areas are highlighted and annotated as "changed from natural features to construction land". This step produces a semantically information-rich change map that intuitively shows the location, extent, and nature of the changed areas, allowing users to clearly understand what changes occurred during the monitoring period. The system outputs change detection results for geographical areas to user terminals for display and storage. The output generates intuitive results showing the changed areas and their semantic information. It receives verification and correction information from users (or domain experts) regarding the change detection results, and uses this information for model training and updates. The entire system forms a self-optimizing cycle, becoming increasingly intelligent and reliable in practical applications. In practice, automatic detection results may contain some errors, thus allowing for manual or other data correction. On one hand, professionals can verify the change results through a visual interface, marking questionable areas to confirm real changes, eliminate false alarms, and supplement missed detections. These manual markings serve as verification and correction feedback to the system. On the other hand, if new data or higher-precision data (such as higher-resolution imagery or ground survey data) becomes available, it can also be used to verify previous detection results. After collecting the aforementioned verification and correction information, the system incorporates it into the next round of processing: for changes requiring immediate correction, the rules or thresholds for change detection can be quickly adjusted based on feedback, resulting in updated results; for changes requiring model improvement, after accumulating a certain amount of correction data, the system triggers model retraining or parameter update processes, enabling the model to learn the corrected patterns; for example, if a certain type of land cover change is frequently falsely detected in the first round, and after multiple expert feedback points, the system will focus on adjusting the feature weights of that type of pattern or adding constraints during retraining to reduce similar errors in the future; through this human-computer interactive feedback mechanism, the system's detection accuracy and robustness will continuously improve, achieving continuous learning and evolution; in long-term monitoring tasks, the feedback mechanism can also help the model adapt to environmental changes (such as new change patterns caused by changes in land use policies), ensuring that the method always maintains a high accuracy rate. At this point, the entire change detection process is closed-loop.

[0021] The technical solution of this invention, arranged in an orderly manner according to control logic, forms an end-to-end change detection process. First, the data acquisition module periodically or on demand acquires multi-temporal remote sensing data, which is then normalized by the preprocessing module and input into the feature extraction module. Subsequently, the extracted multi-temporal features are sent to the temporal fusion module for deep fusion analysis. The fused results are then passed to the change judgment module to generate a change detection conclusion. Finally, the result output module presents the conclusion and enters the feedback loop to continuously improve system performance. Regarding data flow, the data transmission and processing between modules are automatically coordinated by the process control unit. For example, when new image data arrives, feature extraction and change analysis processing are triggered; after processing, result output and storage are triggered, and user feedback is monitored. This collaborative work of a series of modules achieves fully automated processing from raw remote sensing data to change information products, exhibiting excellent scalability and maintainability.

[0022] Based on the foregoing embodiments, the preprocessing of the multi-source remote sensing imagery acquired at different times includes: Aligning multi-source remote sensing image data from different times to the same coordinate system enables geometric correction and registration of multi-source remote sensing image data; For optical images, a radiometric correction algorithm based on an atmospheric radiative transfer model is used to eliminate the effects of atmospheric scattering and absorption, remove cloud and fog masking, and achieve radiometric correction. For SAR images, multi-view processing and filtering denoising operations are performed to reduce speckle noise; All images are processed to achieve a uniform resolution by using bilinear interpolation or nearest neighbor interpolation to adjust the image resolution to a fixed value. For multimodal data, registration and resolution matching are performed between different data sources, and a unified format conversion is carried out.

[0023] Based on the aforementioned embodiments, the specific process for extracting high-level semantic features is as follows: For single-modal images, spatial features are extracted using convolutional neural networks or in combination with a visual Transformer encoder. The specific steps for extracting spatial features using a convolutional neural network combined with a visual Transformer encoder are as follows: Convolutional neural networks (CNNs) with a set number of convolutional stages are constructed based on ResNet-50. After each convolutional stage of a convolutional neural network (CNN), the corresponding feature maps are obtained. A feature pyramid network (FPN) is used to fuse these multi-scale features. High-level features are upsampled and added to low-level features element by element to achieve effective combination of multi-scale features. The output feature map of the last convolutional stage after fusion by the Feature Pyramid Network (FPN) is flattened into a sequence and input into the visual Transformer encoder to output high-level semantic features of a single modality. For multimodal data, a multi-branch feature extraction model or a unified fusion network feature extraction model is used to extract features from different modal images separately or jointly; the specific steps are as follows: Each modality is set up with an independent branch. After each branch extracts features according to the corresponding architecture, a contrastive learning strategy is introduced to calculate the cosine similarity of features of different modalities at the same position. A contrastive loss function is constructed. When the similarity is lower than the set similarity threshold, the network parameters of each branch are updated through backpropagation to make the feature representations of different modalities at the same position more semantically consistent. A hierarchical fusion strategy is adopted, in which modality fusion is performed in the early, middle and late stages of the fusion network. The fused features are then further extracted into high-level semantic features through subsequent convolutional layers, pooling layers and fully connected layers.

[0024] Based on the aforementioned embodiments, deep fusion of spatiotemporal features is achieved, including: For the extracted high-level semantic features at each time point, a Transformer-based multi-head self-attention structure is applied at each spatial location to calculate the correlation weights between high-level semantic features at different times. The high-level semantic features of each frame of the time series image are regarded as nodes in the time graph sequence. Temporally adjacent nodes and spatially adjacent regions are connected by edges to construct a spatiotemporal graph. Through iterative information transmission, the high-level semantic features of adjacent time points and adjacent spatial locations influence each other, realizing the deep fusion of spatiotemporal features. The dimensionality of the fused deep features is reduced, and then normalized so that the deep features are processed at the same scale. Prepare training data with change labels covering different land cover types, change scenarios, and time spans. Input the post-processed data into the model, calculate the error between the model's predicted values ​​and the true labels using the cross-entropy loss function, and optimize the model parameters using stochastic gradient descent or Adam optimization algorithms. Set appropriate learning rates (e.g., 0.001-0.01), batch sizes (e.g., 32-128), and number of iterations (e.g., 100-500). During training, adopt an early stopping strategy: stop training when the loss on the validation set no longer decreases for several consecutive epochs (e.g., 10) to prevent overfitting. Regularly evaluate the model and adjust its hyperparameters based on the evaluation results, such as the number of heads in the multi-head self-attention system and the number of iterations in the graph neural network, to continuously improve the model's performance. Apply the optimized model to actual remote sensing image change detection tasks, outputting information on changed areas in real time, and providing decision support for urban planning, environmental monitoring, and other fields.

[0025] In this scheme, the temporal attention mechanism preserves an independent semantic representation for each time point, and obtains fused features through attention weighting while maintaining complete temporal dimension information. The temporal attention mechanism can automatically focus on significant change regions on the time axis and retain temporal dimension information, while the graph neural network can construct a spatiotemporal graph to achieve deep fusion of spatiotemporal features. The collaboration between the two can capture long-term dependencies and spatial neighborhood consistency, accurately treating regional changes as a related whole.

[0026] Based on the aforementioned embodiments, the step of performing land cover category change analysis based on deep spatiotemporal fusion features, marking areas where land cover category changes have occurred through change analysis, and generating change detection results for geographic areas includes: The decoder employs layer-by-layer upsampling to restore the temporally fused features to the original image resolution. During the decoding process, cross-layer skip connections are used to introduce high-resolution spatial details, and fine semantic segmentation results are output for each time frame. The system performs change analysis on the selected baseline time phase and the current time phase. By comparing the semantic segmentation maps at different times, or by directly outputting the change mask from the model, it determines which areas have experienced changes in land cover categories. It identifies the location range of the changed areas, combines semantic information to provide a category description of the change as a label, and outputs the labeled change results.

[0027] This invention also provides a satellite image land cover change detection system based on long-term time-series information fusion. Please refer to [link to relevant documentation]. Figure 2 ,include: The data acquisition module is used to collect multi-source remote sensing imagery of the same geographical area at different times; to ensure that the acquired multi-temporal images cover the target monitoring period and to ensure the continuity and completeness of the time sequence as much as possible, so as to provide raw data for subsequent processing; The data preprocessing module is used to preprocess multi-source remote sensing image data acquired at different times; The multimodal feature extraction module is used to extract high-level semantic features from preprocessed remote sensing image multi-source data based on a deep learning model. The temporal information fusion module is used to perform deep spatiotemporal fusion of high-level semantic features extracted at each time point in the time dimension to obtain deep spatiotemporal fusion features. Specifically, the correlation weight between high-level semantic features at different times is calculated at each spatial location using a temporal attention mechanism, and the fusion features are obtained through attention weighting. The high-level semantic features of each frame of the time series image are introduced into the graph neural network as nodes in the time graph sequence to model the relationship between temporal features. Here, the graph neural network node represents the high-level semantic feature of a certain image block or pixel position at a specific time. Temporally adjacent nodes and spatially adjacent regions are connected by edges. Through iterative information transmission, the graph neural network enables the high-level semantic features of adjacent times and adjacent spatial locations to influence each other, achieving deep fusion of spatiotemporal features and obtaining deep spatiotemporal fusion features. The change determination module is used to perform change analysis of land cover categories based on deep spatiotemporal fusion features. Through change analysis, it marks areas where land cover categories have changed and generates change detection results for geographic areas. The results output and feedback module is used to output the change detection results of the geographical area to the user terminal for display and storage, and to receive the verification and correction information of the change detection results from the user (or domain expert) for model training and updates.

[0028] Based on the foregoing embodiments, the data preprocessing module includes: The geometric correction submodule is used to align multi-source remote sensing image data from different times to the same coordinate system, thereby achieving geometric correction and registration of multi-source remote sensing image data. The radiometric correction submodule is used to eliminate the effects of atmospheric scattering and absorption, remove cloud and fog masks, and achieve radiometric correction for optical images using a radiometric correction algorithm based on the atmospheric radiative transfer model. The denoising submodule is used for multi-view processing and filtering / denoising operations on SAR images; The unified resolution submodule uses bilinear interpolation or nearest neighbor interpolation to adjust the image resolution to a set fixed value. The format conversion submodule is used to perform registration and resolution matching between different data sources for multimodal data, and to perform unified format conversion.

[0029] Based on the foregoing embodiments, the multimodal feature extraction module includes: The single-modality image feature extraction submodule is used to extract spatial features from single-modality images using a convolutional neural network combined with a visual Transformer encoder, including: CNN Architecture Construction: A Convolutional Neural Network (CNN) architecture with a predetermined number of convolutional stages is constructed based on ResNet-50. For example, a CNN architecture with 5 convolutional stages is set up. The first stage uses a 7x7 convolutional kernel with a stride of 2 and 64 output channels to perform preliminary feature extraction and downsampling on the preprocessed single-modality image. The second stage uses a 3x3 convolutional kernel with a stride of 1 and achieves feature transfer through residual connections, with 128 output channels. The third stage also uses a 3x3 convolutional kernel with a stride of 1 and 256 output channels. The parameters of the fourth stage are the same as the first two stages, with 512 output channels. The fifth stage has 1024 output channels. ResNet-50 is chosen because it introduces residual connections, which can effectively alleviate the gradient vanishing problem during deep network training and is suitable for extracting deep features. The image is input into the ResNet-50 network, and representative deep features are extracted through processing by multiple convolutional layers and pooling layers. The first few layers of the network primarily extract low-level features such as edges and textures. As the network depth increases, the extracted features become increasingly abstract, reflecting higher-level semantic information about ground features. For images of different sizes, they are first divided into fixed-size image blocks (e.g., 256×256 pixels), with each block serving as the network input. After processing, each image block outputs a fixed-dimensional (e.g., 2048-dimensional) feature vector, which effectively represents the semantic information of the corresponding image block. Multi-scale feature extraction and fusion: After each convolutional stage of the convolutional neural network (CNN), the corresponding feature map is obtained. The feature pyramid network (FPN) is used to fuse the multi-scale features, and the high-level features are upsampled and added to the low-level features element by element. The visual Transformer encoder processes the output feature map from the last convolutional stage after fusion by the Feature Pyramid Network (FPN), and inputs it into the visual Transformer encoder. This encoder contains 12 attention heads and 6 coding layers, each consisting of a multi-head self-attention mechanism and a feedforward neural network. The self-attention mechanism captures global spatial relationships by calculating the degree of correlation between elements in the sequence, while the feedforward neural network performs independent nonlinear transformations on each element, outputting high-level semantic features of a single modality. The multimodal feature extraction submodule is used to extract features from different modalities of images using a multi-branch feature extraction model or a unified fusion network feature extraction model. The specific steps are as follows: Branch Architecture Design: Independent branches are set up for each modality. For example, the optical image branch adopts the aforementioned ResNet-50 combined with the visual Transformer architecture, while the SAR image branch adopts an improved convolutional neural network (CNN) architecture. Its first stage uses a 5x5 convolutional kernel with 32 output channels, and subsequent stages use 3x3 kernels with 64, 128, 256, and 512 output channels respectively, to adapt to the characteristics of SAR images. Feature Extraction and Contrastive Learning: After each branch extracts features according to its corresponding architecture, a contrastive learning strategy is introduced to calculate the cosine similarity of features from different modalities at the same location, constructing a contrastive loss function. Where sim is the cosine similarity, τ is the temperature parameter, f1 and f2 are the features of different modes at the same position, and fi is the feature of other i-th point positions; When the similarity is lower than the set similarity threshold (e.g., 0.5), the parameters of each branch network are updated through backpropagation to make the feature representations of different modalities at the same location more semantically consistent. Network architecture design: A hierarchical fusion strategy is adopted to perform modality fusion in the early, middle and late stages of the fusion network. In the early stage, images of different modalities are stitched together by channel and then input into a shared convolutional layer at the input layer. In the middle stage, after each modality has extracted features through some convolutional layers, the features are fused by adding the features. In the late stage, after the features of each modality have been extracted, they are fused through a fully connected layer. Feature extraction and optimization: The fused features are used to extract high-level semantic features through subsequent convolutional layers, pooling layers and fully connected layers. At the same time, a dropout mechanism (dropout rate set to 0.5) is introduced to prevent overfitting, and the network parameters are continuously optimized through backpropagation.

[0030] Based on the aforementioned embodiments, the feature extraction model undergoes self-supervised pre-training, specifically through the following steps: The task of predicting the masked area is to randomly mask a portion of a remote sensing image (e.g., masking 20%), input the masked image into the model, and let the model predict the pixel values ​​and features of the masked area. By calculating the mean squared error loss between the predicted and the true values, the model parameters are updated, enabling the model to learn the local features and contextual relationships of the image. Temporal reconstruction task: Using the location encoding mechanism, timestamp information is converted into location encoding vectors and integrated into image features. Remote sensing images from different time series are shuffled and input into the model, allowing the model to reconstruct the correct temporal order. By calculating temporal prediction loss (such as cross-entropy loss), the model's ability to perceive temporal order is enhanced, and the temporal change characteristics of the images are learned. Cross-modal matching task: Construct images of different modalities, including positive sample pairs (images of different modalities at the same location) and negative sample pairs (images of different modalities at different locations). The model determines whether the input image pairs are positive sample pairs. The model is trained using the cross-entropy loss function, enabling the model to learn the correspondence and semantic association between different modalities. Saving pre-training parameters: After the above pre-training task (e.g., 100 training rounds, batch_size set to 32, learning rate initialized to 0.001 and decaying with each training round), the model parameters are saved to obtain a general remote sensing image representation model. The steps for fine-tuning the model are as follows: Dataset preparation: Select labeled remote sensing image data (e.g., 1000 images) as the fine-tuning dataset, and divide the fine-tuning dataset into a training set (e.g., 80% of the data) and a validation set (e.g., 20% of the data). Model loading and parameter tuning: Load the pre-trained model parameters, freeze some of the lower-level network parameters (e.g., freeze the parameters of the first 5 convolutional layers), and only fine-tune the parameters of the top-level network. For example, set the batch size for fine-tuning to 16, the learning rate to 0.0001, and the training epochs to 30. Model training and evaluation: Input the training set into the model for training, and update the trainable parameters by calculating the classification loss or regression loss (determined according to the specific task); for example, evaluate the model performance (such as accuracy, mean squared error, etc.) on the validation set every 5 training rounds, and adjust the learning rate and training strategy according to the validation set performance to obtain a robust feature extraction model.

[0031] Based on the aforementioned embodiments, the time-series information fusion module includes: The temporal attention mechanism submodule is used to calculate the correlation weights between high-level semantic features at different times by applying a Transformer-based multi-head self-attention structure at each spatial location for the extracted high-level semantic features at each time point. Specifically, the feature vector at each time point is treated as a sequence element, forming a feature sequence. When calculating the attention weights, the feature sequence is first linearly transformed to obtain the query, key, and value matrices. Then, the attention weight matrix is ​​obtained by scaling the dot product attention method, that is, multiplying the transpose of the query and key and dividing by the square root of the feature dimension, and then passing it through the Softmax function. Since a multi-head self-attention structure is used, the feature vector is divided into multiple heads for parallel computation, and finally the results of each head are concatenated to obtain the final attention weights. Through this weight, the model can automatically "pay attention" to those regions that change significantly along the time axis. The higher the weight value, the more important the feature at that time point is in change detection. Unlike traditional methods of compressing time information, this method retains an independent semantic representation for each time point, multiplies the attention weight with the feature vector of the corresponding time point to obtain a weighted feature vector, and then concatenates all the weighted feature vectors to form a fused feature while maintaining complete time dimension information. The spatiotemporal graph construction submodule is used to use the high-level semantic features of each frame of the time series image as nodes in the time graph sequence. Temporally adjacent nodes and spatially adjacent regions are connected by edges. Specifically, a time interval threshold is set. When the time interval between two nodes is less than or equal to the threshold, they are considered to be adjacent nodes and an edge is established. For spatial connections, the spatial coordinates of image blocks or pixel positions are judged. When the spatial distance between the corresponding spatial positions of two nodes is less than the set spatial distance threshold, an edge connection is established to construct the spatiotemporal graph. The spatiotemporal fusion submodule is used to enable high-level semantic features of adjacent time points and spatial locations to influence each other through iterative information transmission. In each iteration, each node collects feature information from its neighboring nodes and updates it in combination with its own features. The update formula is: New node feature = Activation function (Weight matrix × (Self-feature + Sum of features of neighboring nodes) + Bias term), achieving deep fusion of spatiotemporal features. Through multiple iterations (e.g., 10-20 times), deep fusion of spatiotemporal features is achieved, capturing long-term dependencies and spatial neighborhood consistency. For example, it ensures that changes in a city area in adjacent years are regarded as a related whole, rather than scattered point changes. Through the synergistic effect of temporal attention and graph networks, fully fused deep features are output, providing a reliable basis for change determination. The feature post-processing submodule is used to reduce the dimensionality of the fused spatiotemporal features. It employs Principal Component Analysis (PCA) to calculate the covariance matrix of the features, obtain their eigenvalues ​​and eigenvectors, selects the top k eigenvectors with larger eigenvalues, and projects the fused features into the subspace formed by these eigenvectors. After reducing the feature dimensionality, normalization is performed, mapping the dimensionality-reduced eigenvalues ​​to the interval [0,1]. The min-max normalization method is used, i.e., eigenvalue = (eigenvalue - minimum value) / (maximum value - minimum value), so that the depth features are processed at the same scale.

[0032] In this scheme, the temporal attention mechanism preserves an independent semantic representation for each time point, and obtains fused features through attention weighting while maintaining complete temporal dimension information. The temporal attention mechanism can automatically focus on significant change regions on the time axis and retain temporal dimension information, while the graph neural network can construct a spatiotemporal graph to achieve deep fusion of spatiotemporal features. The collaboration between the two can capture long-term dependencies and spatial neighborhood consistency, accurately treating regional changes as a related whole.

[0033] Based on the foregoing embodiments, the change determination module includes: The decoding submodule is used to use the decoder to upsample layer by layer to restore the temporally fused features to the original image resolution. During the decoding process, cross-layer skip connections are used to introduce high-resolution spatial details and output fine semantic segmentation results for each time frame. The decision submodule is used to perform change analysis on the selected reference time phase and the current time phase. By comparing the semantic segmentation maps at different times, or by directly outputting the change mask from the model, it determines which areas have changed the land cover categories. It identifies the location range of the changed areas, combines semantic information to give a category description of the change as a label, and outputs the labeled change results.

[0034] The change determination module in this scheme generates specific change detection results based on the fused spatiotemporal features. It can be considered the decoding and decision-making part of the model. On one hand, the temporally fused features are upsampled layer by layer by the decoder to restore the original image resolution, and a detailed semantic segmentation result (i.e., a land cover classification map at each time point) is output for each time frame. During the decoding process, cross-layer skip connections are used to introduce high-resolution spatial details, improving the accuracy of the segmentation results. On the other hand, the system performs change analysis on the selected baseline time phase and the current time phase (or the entire time range): by comparing the semantic segmentation maps at different times, or directly using the change mask output by the model, it determines which areas have experienced changes in land cover categories. The change determination module identifies the location range of the changed areas and provides a category description of the change based on semantic information. For example, if an area changes from the "vegetation" category to "built land," the output change result marks that the area has undergone a change from vegetation to urban construction. In terms of design, two methods can be used to obtain semantically meaningful change patches: one is post-classification comparison (performing difference operations on the segmentation results of each time phase to obtain the changed regions and category changes); the other is direct change classification (training the model to directly output a change detection map, where the value of each pixel represents the change type or whether a change has occurred). The implementation scheme of this invention tends to combine the advantages of both methods: based on deep feature fusion of the model, it outputs the semantic segmentation results of each time phase, and then the change determination module compares the baseline time phase with the current time phase to obtain the changed regions and assigns them semantic labels. This approach ensures that the change detection results have both localization accuracy and semantic interpretability.

[0035] To more intuitively illustrate the operational logic of the system and method of this invention, the following example of urban expansion monitoring will be used to explain the collaborative work process of each module and its advantages. Suppose we need to monitor the land use changes in a city over the past ten years, including the expansion of urban building land.

[0036] First, the data acquisition module will collect satellite remote sensing imagery data for the city for each of the past ten years. For example, from 2015 to 2025, one cloud-free, high-resolution optical satellite image will be selected each year to form a time-series image set. If the optical data for some years is severely affected by clouds, SAR radar imagery from the same period can be introduced as a supplement, achieving cross-modal data acquisition. Through this collection of multi-source data, surface observation imagery covering a ten-year period is ensured, providing a complete chain of evidence for detecting urban expansion.

[0037] Next, the data preprocessing module processes each of the ten image periods. The system registers all years' images to the same urban geographic coordinate system, ensuring accurate alignment of the same locations within the city across different years' images. For example, the pixel grids of a new urban area can be matched one-to-one across images from different years. Then, radiometric correction and brightness normalization are performed on the images to correct for overall color variations caused by seasonal and lighting differences. Radar images (if any) are also filtered, normalized, and resampled to the same resolution as the optical images. Furthermore, the preprocessing module detects cloud cover or shadow areas in each image: if a 2018 image is partially obscured by clouds, the system replaces that area with the corresponding 2018 radar image or a clear optical image from a nearby date, ensuring the data for each year is as complete and reliable as possible. After preprocessing, ten high-quality aligned annual remote sensing images are obtained, accurately reflecting the evolution of land cover in the city over the past decade.

[0038] Then, these images are processed by the feature extraction module for semantic information extraction. The system loads a pre-trained deep learning model to perform semantic segmentation prediction on the images for each year. The model classifies pixels in urban areas into various land cover categories, such as "built-up land," "roads," "vegetation," "water bodies," and "bare land." Taking images from earlier years (such as 2015) as an example, the model may identify that the urban built-up area is mainly concentrated in the old city area, while the suburbs are mostly marked as vegetation or farmland. In subsequent years, comparing images from years such as 2020 and 2025, the model can find that areas that were originally vegetated in the suburbs are gradually being covered by new buildings, and those areas are correspondingly changed from the "vegetation" category to the "built-up land" category in the semantic segmentation results. Because the model's feature extraction process combines self-supervised pre-training experience and multimodal data fusion capabilities, it can accurately classify land cover on each image. Even if the image quality of a certain year is slightly inferior or contains some noise, the model can maintain a relatively stable judgment based on features learned from other years. The semantic classification map generated in this step lays the foundation for subsequent steps.

[0039] The temporal information fusion module then jointly analyzes the features and initial classification results of the aforementioned annual images, focusing on capturing patterns of urban land use change. This module plays a crucial role in this embodiment: utilizing a temporal attention mechanism and a spatiotemporal graph neural network, the model longitudinally compares the change trajectory of each pixel category from 2015 to 2025. For areas that remain unchanged for many years (such as old urban areas that have always been urban buildings or nature reserves that have always been forests), the model's temporal attention assigns them a lower change attention, representing a "stable state"; while for areas on the outskirts of the suburbs that show signs of development year by year, the model's attention weight gradually increases, indicating that it has "perceived" the abnormal changes in these locations over time. For example, a piece of land was classified as "vegetation" from 2015 to 2018, some pixels became "bare land" starting in 2019 (possibly indicating the start of construction), and by 2021 most of it had become "building land." The model will integrate the feature changes over these years to determine that a continuous semantic shift from vegetation to buildings has occurred here. Spatiotemporal graph neural networks further ensure spatial consistency: if the development of a new area involves a large, continuous region, the graph network connects all the nodes in the area, allowing the model to recognize through information propagation that this is a related, holistic change, rather than just isolated pixel changes. This avoids fragmented detection results, instead labeling the expansion of the new area as a coherent change.

[0040] The change determination module generates urban expansion detection output based on the results of the fusion analysis. For urban expansion, the focus is on the increase in built-up land. Therefore, the system performs a pixel-level comparison between the classification map of the baseline year (e.g., 2015) and the classification map of the latest year (2025). With the support of the fusion module, these classification maps have been optimized for high accuracy. The system identifies areas where the pixel category changed from "non-building" (e.g., vegetation, bare land) in 2015 to "built-up land" in 2025, and aggregates these pixels into contiguous change patches. These patches correspond to the new expansion of the urban built-up area over the past decade. Each change patch includes semantic information ("newly built-up land") and its boundaries. For example, the system might output a polygonal vector layer outlining the newly added development area in the southeast of the city, indicating that it was formerly primarily farmland and is now a residential area. Simultaneously, if areas show buildings being converted to other uses (e.g., some buildings being demolished and turned into green space), the system can also detect and mark them. However, in the context of urban expansion, such "negative growth" is often rare or localized, and the system's main output will focus on newly built areas. The output can quantify the area of ​​urban expansion, summarizing how many square kilometers of urban construction land were added between 2015 and 2025, and indicating the specific locations.

[0041] Finally, the results output and feedback module provides the above change information to urban planning managers, allowing them to verify and provide feedback. The output change map can be overlaid on city maps or remote sensing imagery, highlighting the newly expanded urban area with a high-contrast color. This helps decision-makers intuitively understand the direction and scale of urban sprawl over a decade. If managers find discrepancies between the detection results and the actual situation—for example, an area marked as expansion is actually a temporary construction site that has not yet formed a permanent building—they can mark the area as a "false positive" through the system interface. After receiving feedback, the system will retrieve the characteristics of that area from the database for each year and find that the original construction site returned to vacancy in the last year (assuming that construction was actually halted after completion). The reason for the misjudgment as building expansion was that short-term changes misled the model. With the feedback and annotations from managers, the system will adjust the model's judgment strategy for similar short-term changes during the next training, potentially requiring areas that have maintained changes for many consecutive years to be judged as permanent changes, thereby avoiding future misjudgments of such temporary changes as trend changes. For other newly established urban areas that have been verified, the feedback from administrators is equivalent to approval of the model's output. The system also records these "correctly detected" instances, continuously accumulating positive and negative examples to strengthen the model's discriminative ability. Through the above closed-loop process, in real-world scenarios, the system of this invention can continuously calibrate itself to adapt to different urban expansion patterns and data characteristics, ensuring accurate and reliable long-term monitoring.

[0042] In summary, the urban expansion monitoring example demonstrates how this system can successfully capture gradual land use changes using long-term remote sensing data. Compared to traditional methods that rely on comparing two consecutive images, this invention, by fusing information from a complete ten-year sequence, can identify small, cumulative changes over the years and recognize them as significant changes, thus providing more comprehensive and detailed change intelligence. For example, the process of a suburban area developing from scattered buildings to a gradually contiguous area may not be obvious in a comparison of any two adjacent years, but through long-term analysis using this method, the overall expansion trend is clearly visible. This is of great value to urban planning departments in predicting urban development in advance and rationally allocating infrastructure. Simultaneously, the use of multimodal data ensures that data is available regardless of rainy weather or other sensor limitations, preventing monitoring interruptions; the addition of a feedback mechanism ensures that the results can be continuously corrected and improved to meet the accuracy requirements of actual management. This embodiment confirms the effectiveness and superiority of the system design and methodology of this invention.

[0043] The semantic change detection system and method based on long-term remote sensing imagery proposed in this invention integrates multi-source information and innovative algorithms, and has the following outstanding advantages in practical applications: By utilizing long-term time series analysis, the impact of single-period errors is reduced: By comprehensively analyzing the data redundancy of multiple remote sensing images, this method effectively reduces the interference of poor image quality in any single period on the results. Even if data at certain times contains cloud shadows, noise, or other anomalies, long-term time series analysis can compensate from information from preceding and following times, avoiding the conclusion of change based on a single erroneous observation. This makes the change detection results more robust, significantly reducing false positives and false negatives, and ensuring a more reliable grasp of true changes.

[0044] Sensitive Capture of Subtle, Gradual Changes: Traditional two-phase methods often fail to detect slowly occurring changes. This invention, through a time attention mechanism, fully leverages information from the time dimension, enhancing sensitivity to slow and subtle changes. For example, the gradual degradation of vegetation cover and the annual expansion of urban edges accumulate into significant change signals in the model. Therefore, this method can provide more refined change trajectories than traditional methods when dealing with long-term monitoring tasks, and can also support the analysis and prediction of change trends.

[0045] By introducing graph neural networks to ensure spatial context consistency: By constructing a spatiotemporal graph and using graph neural networks to propagate features, this method incorporates spatial neighborhood information and temporal coherence information into change detection. This results in more coherent and complete change detection results. Change areas are output as patches rather than scattered, disorganized sets of pixels; simultaneously, spatially related changes (e.g., concurrent changes in adjacent plots) can be uniformly identified. This spatial consistency reduces false change noise, which is highly beneficial for improving the accurate localization and shape characterization of change patches.

[0046] Multimodal contrastive learning enhances generalization ability: This system supports the fusion of multi-source remote sensing data such as optical imagery and SAR radar imagery, and aligns the feature representations of different modalities through cross-modal contrastive learning. This enables the model to maintain performance under various data conditions: when optical imagery is limited (e.g., during cloudy seasons), the addition of radar data still ensures the continuity of detection; the model learns the consistency of the same ground features under different sensors during training, thus demonstrating good adaptability to new areas and new sensors in practical applications. Compared to methods that rely solely on a single modality, this invention is more robust to changes in data sources and has a wider range of applications.

[0047] Self-supervised pre-training and feedback mechanisms improve accuracy: Through self-supervised feature extraction pre-training, the model fully utilizes massive amounts of unlabeled data to mine general features, enabling it to achieve high initial accuracy even with limited labeled samples. This training strategy also enhances the model's transfer and generalization capabilities to unknown scenarios. Furthermore, the system's human-computer interaction feedback loop allows the model to continuously learn after deployment: whenever the user corrects the results, the model iteratively updates, gradually correcting its shortcomings. In the long run, the system's accuracy will increase with use, maintaining high reliability in different application scenarios. This continuous optimization capability is not present in traditional fixed models, ensuring that the method of this invention remains accurate and robust in complex and ever-changing real-world environments.

[0048] In summary, the system and method provided by this invention successfully overcome many limitations of traditional change detection methods through deep fusion of long-term, multi-source data and the synergistic effect of innovative attention mechanisms, graph neural networks, and self-supervised learning technologies. It can automatically and efficiently detect semantic changes in land cover, possessing both high accuracy and strong robustness. It can be widely applied in fields such as land resource monitoring, ecological environment change tracking, and urban development supervision, providing strong support for relevant decision-making and demonstrating significant practical value and promising prospects for wider application.

[0049] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting ground feature changes in satellite images based on long-term time-series information fusion, characterized in that, include: Collect multi-source remote sensing imagery data of the same geographic area at different times; Preprocessing is performed on the multi-source remote sensing imagery acquired at different times; Based on a deep learning model, high-level semantic features are extracted from preprocessed remote sensing image multi-source data. The high-level semantic features extracted at each time point are deeply spatiotemporally fused along the time dimension to obtain deep spatiotemporally fused features, including: The correlation weights between high-level semantic features at different times are calculated at each spatial location using a temporal attention mechanism, and the fused features are obtained through attention weighting. The high-level semantic features of each frame of the time series image are introduced into the graph neural network as nodes in the time graph sequence to model the relationship of the temporal features. The graph neural network node represents the high-level semantic features of a certain image block or pixel position at a specific time. Temporally adjacent nodes and spatially adjacent regions are connected by edges. Through iterative information transmission, the graph neural network enables the high-level semantic features of adjacent times and adjacent spatial positions to influence each other, achieving deep fusion of spatiotemporal features and obtaining deep spatiotemporal fusion features. Based on the deep spatiotemporal fusion characteristics, we perform land cover category change analysis, mark the areas where land cover category changes have occurred through change analysis, and generate change detection results for geographic areas. The system outputs the change detection results of the geographical area to the user terminal for display and storage, receives the user's verification and correction information on the change detection results, and uses the verification and correction information for model training and updates.

2. The satellite image land cover change detection method based on long-term time-series information fusion as described in claim 1, characterized in that, The preprocessing of multi-source remote sensing imagery acquired at different times includes: Aligning multi-source remote sensing image data from different times to the same coordinate system enables geometric correction and registration of multi-source remote sensing image data; For optical images, a radiometric correction algorithm based on an atmospheric radiative transfer model is used to eliminate the effects of atmospheric scattering and absorption, remove cloud and fog masking, and achieve radiometric correction. Perform multi-view processing and filtering / denoising operations on SAR images; The image resolution is adjusted to a fixed value using bilinear interpolation or nearest neighbor interpolation. For multimodal data, registration and resolution matching are performed between different data sources, and a unified format conversion is carried out.

3. The satellite image land cover change detection method based on long-term time-series information fusion as described in claim 1, characterized in that, The specific process for extracting high-level semantic features is as follows: For single-modal images, a convolutional neural network combined with a visual Transformer encoder is used to extract spatial features, including: Convolutional neural networks with a predetermined number of convolutional stages are constructed based on ResNet-50; After each convolutional stage of the convolutional neural network, the corresponding feature map is obtained, and a feature pyramid network is used to fuse multi-scale features. The high-level features are upsampled and added to the low-level features element by element. The output feature map of the last convolutional stage after the feature pyramid network is fused is input into the visual Transformer encoder to output high-level semantic features of a single modality. For multimodal data, a multi-branch feature extraction model or a unified fusion network feature extraction model is used to extract features from different modalities of the image, including: Each modality is set up with an independent branch. After each branch extracts features according to the corresponding architecture, a contrastive learning strategy is introduced to calculate the cosine similarity of features of different modalities at the same position. A contrastive loss function is constructed. When the similarity is lower than the set similarity threshold, the network parameters of each branch are updated through backpropagation to make the feature representations of different modalities at the same position more semantically consistent. A hierarchical fusion strategy is adopted to perform modality fusion in the early, middle and late stages of the fusion network. The fused features are then used to extract high-level semantic features through subsequent convolutional layers, pooling layers and fully connected layers.

4. The satellite image land cover change detection method based on long-term time-series information fusion as described in claim 1, characterized in that, The deep fusion that achieves spatiotemporal features includes: For the extracted high-level semantic features at each time point, a Transformer-based multi-head self-attention structure is applied at each spatial location to calculate the correlation weights between high-level semantic features at different times. The high-level semantic features of each frame of the time series are regarded as nodes in the time graph sequence. Temporally adjacent nodes and spatially adjacent regions are connected by edges to construct a spatiotemporal graph. Through iterative information transmission, high-level semantic features of adjacent time points and adjacent spatial locations influence each other, achieving deep fusion of spatiotemporal features; The fused spatiotemporal features are subjected to dimensionality reduction and normalization.

5. The satellite image land cover change detection method based on long-term time-series information fusion as described in claim 1, characterized in that, The method involves analyzing land cover category changes based on deep spatiotemporal fusion features, marking areas where land cover category changes have occurred through change analysis, and generating change detection results for geographic areas, including: The decoder employs layer-by-layer upsampling to restore the temporally fused features to the original image resolution. During the decoding process, cross-layer skip connections are used to introduce high-resolution spatial details, and fine semantic segmentation results are output for each time frame. The system performs change analysis on the selected baseline time phase and the current time phase. By comparing the semantic segmentation maps at different times, or by directly outputting the change mask from the model, it determines which areas have experienced changes in land cover categories. It identifies the location range of the changed areas, combines semantic information to provide a category description of the change as a label, and outputs the labeled change results.

6. A satellite imagery-based ground feature change detection system based on long-term time-series information fusion, characterized in that, include: The data acquisition module is used to collect multi-source remote sensing imagery data of the same geographical area at different times; The data preprocessing module is used to preprocess multi-source remote sensing image data acquired at different times; The multimodal feature extraction module is used to extract high-level semantic features from preprocessed remote sensing image multi-source data based on a deep learning model. The temporal information fusion module is responsible for performing deep spatiotemporal fusion of the high-level semantic features extracted from each time point in the time dimension to obtain deep spatiotemporal fused features, including: The correlation weights between high-level semantic features at different times are calculated at each spatial location using a temporal attention mechanism, and the fused features are obtained through attention weighting. The high-level semantic features of each frame of the time series image are introduced into the graph neural network as nodes in the time graph sequence to model the relationship of the temporal features. The graph neural network node represents the high-level semantic features of a certain image block or pixel position at a specific time. Temporally adjacent nodes and spatially adjacent regions are connected by edges. Through iterative information transmission, the graph neural network enables the high-level semantic features of adjacent times and adjacent spatial positions to influence each other, achieving deep fusion of spatiotemporal features and obtaining deep spatiotemporal fusion features. The change determination module is used to perform change analysis of land cover categories based on deep spatiotemporal fusion features. Through change analysis, it marks areas where land cover categories have changed and generates change detection results for geographic areas. The results output and feedback module is used to output the change detection results of the geographical area to the user terminal for display and storage, receive the user's verification and correction information on the change detection results, and use the verification and correction information for model training and updates.

7. The satellite image land cover change detection system based on long-term time-series information fusion as described in claim 6, characterized in that, The data preprocessing module includes: The geometric correction submodule is used to align multi-source remote sensing image data from different times to the same coordinate system, thereby achieving geometric correction and registration of multi-source remote sensing image data. The radiometric correction submodule is used to eliminate the effects of atmospheric scattering and absorption, remove cloud and fog masks, and achieve radiometric correction for optical images using a radiometric correction algorithm based on the atmospheric radiative transfer model. The denoising submodule is used for multi-view processing and filtering / denoising operations on SAR images; The unified resolution submodule is used to adjust the image resolution to a set fixed value using bilinear interpolation or nearest neighbor interpolation. The format conversion submodule is used to perform registration and resolution matching between different data sources for multimodal data, and to perform unified format conversion.

8. The satellite image land cover change detection system based on long-term time-series information fusion as described in claim 6, characterized in that, The multimodal feature extraction module includes: The single-modality image feature extraction submodule is used to extract spatial features from single-modality images using a convolutional neural network combined with a visual Transformer encoder, including: Convolutional neural networks with a predetermined number of convolutional stages are constructed based on ResNet-50; After each convolutional stage of the convolutional neural network, the corresponding feature map is obtained, and a feature pyramid network is used to fuse multi-scale features. The high-level features are upsampled and added to the low-level features element by element. The output feature map of the last convolutional stage after the feature pyramid network is fused is input into the visual Transformer encoder to output high-level semantic features of a single modality. The multimodal feature extraction submodule is used to extract features from different modalities of images using a multi-branch feature extraction model or a unified fusion network feature extraction model, including: Each modality is set up with an independent branch. After each branch extracts features according to the corresponding architecture, a contrastive learning strategy is introduced to calculate the cosine similarity of features of different modalities at the same position. A contrastive loss function is constructed. When the similarity is lower than the set similarity threshold, the network parameters of each branch are updated through backpropagation to make the feature representations of different modalities at the same position more semantically consistent. A hierarchical fusion strategy is adopted to perform modality fusion in the early, middle and late stages of the fusion network. The fused features are then used to extract high-level semantic features through subsequent convolutional layers, pooling layers and fully connected layers.

9. The satellite image land cover change detection system based on long-term time-series information fusion as described in claim 6, characterized in that, The time-series information fusion module includes: The temporal attention mechanism processing submodule is used to apply a Transformer-based multi-head self-attention structure to calculate the correlation weights between high-level semantic features at different times for the extracted high-level semantic features at each spatial location. The spatiotemporal graph construction submodule is used to treat the high-level semantic features of each frame of the time series as nodes in the time graph sequence. Temporally adjacent nodes and spatially adjacent regions are connected by edges to construct the spatiotemporal graph. The spatiotemporal fusion submodule is used to achieve deep fusion of spatiotemporal features by enabling high-level semantic features of adjacent time points and adjacent spatial locations to influence each other through iterative information transmission. The feature post-processing submodule is used to perform dimensionality reduction and normalization on the fused spatiotemporal features.

10. The satellite image land cover change detection system based on long-term time-series information fusion as described in claim 6, characterized in that, The change determination module includes: The decoding submodule is used to use the decoder to upsample layer by layer to restore the temporally fused features to the original image resolution. During the decoding process, cross-layer skip connections are used to introduce high-resolution spatial details and output fine semantic segmentation results for each time frame. The decision submodule is used to perform change analysis on the selected reference time phase and the current time phase. By comparing the semantic segmentation maps at different times, or by directly outputting the change mask from the model, it determines which areas have changed the land cover categories. It identifies the location range of the changed areas, combines semantic information to give a category description of the change as a label, and outputs the labeled change results.

Images