Remote sensing image change detection method and device

Abstract

The invention provides a remote sensing image change detection method and device, which can be applied to the technical field of multi-modal artificial intelligence and remote sensing image analysis. The remote sensing image change detection method comprises the following steps: acquiring a multi-temporal remote sensing image of the same imaging target, and setting a task description language for representing the multi-modal processing requirement of the multi-temporal remote sensing image; performing language coding processing on the task description language to obtain a semantic vector, and performing feature extraction on the multi-temporal remote sensing image according to the semantic vector to obtain a unified visual feature; pre-processing the unified visual features by using the trained visual feature adapter to obtain pre-processed visual features, and executing task reasoning on the pre-processed visual features by using the trained thinking chain model according to the semantic vector to obtain a thinking chain task reasoning result; and performing change detection processing on the multi-temporal remote sensing image based on the thinking chain task reasoning result to obtain a change detection result of the multi-temporal remote sensing image.

Description

Technical Field

[0001] This invention relates to the field of multimodal artificial intelligence and remote sensing image analysis technology, and specifically to a method and apparatus for detecting changes in remote sensing images. Background Technology

[0002] Remote sensing image change detection and inference technology identifies land cover change information by comparing remote sensing images from different time periods, and has irreplaceable value in fields such as natural resource monitoring, disaster early warning, and land planning. Against the backdrop of accelerating global urbanization, this technology can quantitatively analyze environmental evolution such as construction land expansion, deforestation, and water pollution.

[0003] Traditional methods mainly include threshold segmentation, principal component analysis, and post-classification comparison. Pixel-based change detection is susceptible to radiation differences and noise interference, and visual interpretation relies on expert experience and is inefficient. Statistical models such as Markov random fields have limited ability to represent complex scenes and are difficult to adapt to the needs of multi-temporal and multi-sensor data fusion. Although deep learning models (such as U-Net and Siamese CNN) are superior to traditional algorithms in remote sensing change detection, existing technologies have the following limitations: (1) Task isolation: Traditional methods independently perform tasks such as change detection and target recognition, lacking cross-task knowledge transfer; (2) Semantic fragmentation: Visual features and semantic understanding are separated, making it impossible to establish a complete cognitive chain of "pixel change → object change → semantic conversion"; (3) Static processing: Fixed processing flow is difficult to adapt to the reasoning needs of different scenarios (such as building demolition and forest fire requiring different analysis paths).

[0004] Therefore, under the aforementioned technological background, remote sensing image change understanding is of great research significance. Remote sensing image change understanding not only needs to detect changed areas and identify change types, but also needs to achieve deep reasoning about complex changes through the mutual coordination and verification of change detection, semantic segmentation, target detection, and change description. Summary of the Invention

[0005] In view of the above problems, the present invention provides a method and apparatus for detecting changes in remote sensing images, which is used to solve at least one of the above technical problems.

[0006] According to a first aspect of the present invention, a method for detecting changes in remote sensing images is provided, comprising: acquiring multi-temporal remote sensing images of the same imaging target, and setting a task description language for characterizing the multi-modal processing required for multi-temporal remote sensing images; performing language encoding processing on the task description language to obtain semantic vectors, and extracting features from the multi-temporal remote sensing images based on the semantic vectors to obtain unified visual features; preprocessing the unified visual features using a trained visual feature adapter to obtain preprocessed visual features, and performing task inference on the preprocessed visual features based on the semantic vectors using a trained thought chain model to obtain a thought chain task inference result; and performing change detection processing on the multi-temporal remote sensing images based on the thought chain task inference result to obtain a change detection result for the multi-temporal remote sensing images.

[0007] According to an embodiment of the present invention, the acquisition of multi-temporal remote sensing images of the same imaging target includes: acquiring a first remote sensing image of the same imaging target at a first time and a second remote sensing image at a second time; performing radiometric correction and registration processing on the first remote sensing image to obtain a preprocessed first remote sensing image, and performing radiometric correction and registration processing on the second remote sensing image to obtain a preprocessed second remote sensing image.

[0008] According to an embodiment of the present invention, the above-mentioned language encoding processing of the task description language to obtain semantic vectors includes: performing word vector embedding fusion operation, bidirectional Transformer encoding operation, and pooling operation on words related to land features in the task description language to obtain semantic vectors; wherein, based on the semantic vectors, feature extraction of multi-temporal remote sensing images to obtain unified visual features includes: using a pre-trained visual-language large model to perform cross-modal feature alignment operation based on contrastive learning and mapping operation based on the same semantic space on the first remote sensing image and the second remote sensing image to obtain unified visual features, wherein the same semantic space is represented by the semantic vectors.

[0009] According to an embodiment of the present invention, the trained visual feature adapter and the trained thought chain inference model are obtained through the following operations: using the visual feature adapter to perform super-resolution image reconstruction on the visual feature samples of multi-temporal remote sensing image samples to obtain preprocessed visual feature samples; pre-training the visual feature adapter and multiple sub-task models for multi-temporal remote sensing image processing to obtain a pre-trained visual feature adapter and multiple pre-trained sub-task models; iteratively training the thought chain inference model and the pre-trained visual feature adapter using semantic vector samples corresponding to the visual feature samples, the preprocessed visual feature samples, and the multiple pre-trained sub-task models to obtain a trained visual feature adapter and a trained thought chain inference model.

[0010] According to an embodiment of the present invention, the above-mentioned pre-training of the visual feature adapter and multiple sub-task models for multi-temporal remote sensing image processing to obtain a pre-trained visual feature adapter and multiple pre-trained sub-task models includes: using the visual feature adapter with randomly initialized parameters to perform channel-dimensional concatenation processing on the multi-temporal visual features of the multi-temporal remote sensing image samples to obtain multi-temporal feature map samples; processing the multi-temporal feature map samples using multiple sub-task models to obtain predicted values ​​representing different processing tasks of the multi-temporal remote sensing image samples; calculating the loss between the predicted values ​​and the true values ​​of the multi-temporal remote sensing image samples to obtain pre-trained loss values; and updating the parameters of the visual feature adapter and multiple sub-task models using the pre-trained loss values ​​through backpropagation to obtain the pre-trained visual feature adapter and multiple pre-trained sub-task models.

[0011] According to embodiments of the present invention, the above-mentioned sub-task models include a change detection task model, a change description task model, a target recognition task model, and a semantic segmentation task model.

[0012] According to an embodiment of the present invention, the above-mentioned iterative training of the thought chain inference model and the pre-trained visual feature adapter using semantic vector samples corresponding to visual feature samples, pre-processed visual feature samples, and multiple pre-trained sub-task models to obtain the trained visual feature adapter and the trained thought chain inference model includes: in the first iteration, based on the semantic vector samples, the thought chain model is used to perform task decomposition, and based on the result of task decomposition, all sub-task models are called to perform processing on multi-temporal remote sensing image samples; in other iterations, based on the semantic vector samples, the pre-processed visual feature samples and multiple pre-trained sub-task models are used to iteratively perform task selection and performance verification operations based on reinforcement learning algorithms on the thought chain model and the pre-trained visual feature adaptation to obtain the trained visual feature adapter and the trained thought chain inference model.

[0013] According to an embodiment of the present invention, in the above-mentioned other iteration processes, based on semantic vector samples, using preprocessed visual feature samples and multiple pre-trained sub-task models, iteratively performing task selection and performance verification operations based on reinforcement learning algorithms on the mental chain model and the pre-trained visual feature adaptation to obtain a trained visual feature adapter and a trained mental chain inference model includes: based on semantic vector samples, performing task inference operations on preprocessed visual feature samples using the mental chain model to obtain task inference results; selecting a pre-trained target task model from multiple pre-trained sub-task models based on the task inference results, and using the pre-trained target task model to perform image processing on multi-temporal remote sensing image samples to obtain intermediate results; performing verification operations on the intermediate processing results, and performing parameter update operations on the pre-trained visual feature adapter and mental chain inference model based on the verification results to obtain a trained visual feature adapter and a trained mental chain inference model.

[0014] According to an embodiment of the present invention, the above-mentioned selection of a pre-trained target task model from multiple pre-trained sub-task models based on task inference results includes: determining the current state in the current training round based on task inference results and the output results of the sub-task models in the previous training round; determining the selection action based on the current state and the value function; and selecting the pre-trained target task model from multiple pre-trained sub-task models based on the selection action.

[0015] According to a second aspect of the present invention, a remote sensing image processing apparatus is provided, comprising: an image acquisition and task setting module, configured to acquire multi-temporal remote sensing images of the same imaging target and set a task description language for characterizing the multi-modal processing required for multi-temporal remote sensing images; a visual feature acquisition module, configured to perform language encoding processing on the task description language to obtain semantic vectors, and extract features from the multi-temporal remote sensing images based on the semantic vectors to obtain unified visual features; a thought chain reasoning module, configured to preprocess the unified visual features using a trained visual feature adapter to obtain preprocessed visual features, and perform task reasoning on the preprocessed visual features based on the semantic vectors using a trained thought chain model to obtain a thought chain task reasoning result; and a multi-temporal remote sensing image processing module, configured to perform image processing on the multi-temporal remote sensing images based on the thought chain task reasoning result to obtain the processing result of the multi-temporal remote sensing images.

[0016] A third aspect of the present invention provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0017] A fourth aspect of the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions, when executed by a processor, implement the steps of the above-described method.

[0018] A fifth aspect of the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0019] The remote sensing image change detection method provided by this invention aligns textual information with visual features through semantic vectors encoded by a task description language, effectively solving the false detection problem caused by radiometric differences and registration errors in traditional methods. It achieves multimodal collaborative enhancement in the change detection and inference process of multi-temporal remote sensing images. Simultaneously, by utilizing a trained thought chain model through step-by-step inference, it can distinguish between real changes and interference factors in complex scenes and dynamically allocate computing resources according to task requirements, improving processing efficiency and reducing computational costs. Furthermore, the remote sensing image change detection method provided by this invention displays the collaborative inference process of multiple tasks through the thought chain, meeting users' needs for interpretability, greatly improving the effect and efficiency of remote sensing image processing, and providing a more convenient and reliable service for practical applications. Attached Figure Description

[0020] The above-described features, other objects, and advantages of the present invention will become clearer from the following description of embodiments of the invention with reference to the accompanying drawings, in which:

[0021] Figure 1 An application scenario diagram of the remote sensing image change detection method according to an embodiment of the present invention is shown.

[0022] Figure 2 A flowchart of a remote sensing image change detection method according to an embodiment of the present invention is shown.

[0023] Figure 3 A training and testing framework diagram for remote sensing image change detection and understanding based on thought chain reasoning according to an embodiment of the present invention is shown.

[0024] Figure 4 A structural block diagram of a remote sensing image change detection device according to an embodiment of the present invention is shown.

[0025] Figure 5 A block diagram of an electronic device suitable for implementing a remote sensing image change detection method according to an embodiment of the present invention is shown. Detailed Implementation

[0026] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the invention for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0027] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0028] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0029] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0030] To address the problems existing in the detection and reasoning of changes in remote sensing images in the prior art, this invention provides a method and apparatus for detecting changes in remote sensing images. By using a thought chain reasoning technique, the detection and reasoning of changes in remote sensing images is decomposed into a series of interrelated sub-tasks such as change detection, semantic segmentation, image description, and target recognition. Through the mutual coordination and verification of these sub-tasks, a deep understanding of the complex changes in remote sensing images is achieved.

[0031] Figure 1 An application scenario diagram of the remote sensing image change detection method according to an embodiment of the present invention is shown.

[0032] like Figure 1 As shown, application scenario 100 according to this embodiment may include scenarios such as multimodal artificial intelligence and remote sensing image analysis. Network 104 is used as a medium to provide a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. Network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0033] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0034] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0035] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0036] It should be noted that the remote sensing image change detection method provided in this embodiment of the invention can generally be executed by server 105. Correspondingly, the remote sensing image change detection device provided in this embodiment of the invention can generally be located in server 105. The remote sensing image change detection method provided in this embodiment of the invention can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the remote sensing image change detection device provided in this embodiment of the invention can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0037] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0038] The following will be based on Figure 1 The described scene, through Figures 2-3 The remote sensing image change detection method of the disclosed embodiments is described in detail.

[0039] Figure 2 A flowchart of a remote sensing image change detection method according to an embodiment of the present invention is shown.

[0040] like Figure 2 As shown, the remote sensing image change detection method of this embodiment includes operations S210 to S240.

[0041] In operation S210, multi-temporal remote sensing images of the same imaging target are acquired, and a task description language is set to characterize the multi-modal processing required for multi-temporal remote sensing images.

[0042] For example, remote sensing images of the same region at time t1 and time t2 are acquired, and the multi-temporal remote sensing images of the same region are preprocessed to complete the radiometric correction and registration of the multi-temporal remote sensing images.

[0043] Task description language, such as "detect changes in buildings in multi-temporal remote sensing images".

[0044] In operation S220, the task description language is encoded to obtain semantic vectors, and features are extracted from multi-temporal remote sensing images based on the semantic vectors to obtain unified visual features.

[0045] Feature extraction can be performed on multi-temporal remote sensing images, for example, using a pre-trained visual-language large model. The visual-language large model is a multimodal large model used to fuse the semantic information of the remote sensing images at time t1, time t2, and the task description language to generate visual features containing high-level semantics and change information.

[0046] In operation S230, the trained visual feature adapter is used to preprocess the unified visual features to obtain the preprocessed visual features. Based on the semantic vector, the trained thought chain model is used to perform task inference on the preprocessed visual features to obtain the thought chain task inference result.

[0047] In the process of reasoning for the thought chain task, the thought chain reasoning model first needs to be trained. The trained thought chain reasoning model is then used to reason about the change detection task of multi-temporal remote sensing images. The reasoning results are used to select the processing task of multi-temporal remote sensing images to be performed, such as semantic segmentation and object detection of multi-temporal remote sensing images.

[0048] For example, unified visual features are input into the thought chain inference model through a visual feature adapter. The thought chain inference model decomposes the image change understanding task into one or more sub-tasks and selectively calls at least one preset downstream visual network to execute the corresponding sub-task, generating intermediate results. The intermediate results are then subjected to performance verification to determine whether the inference process meets the preset termination conditions. If the determination result is negative, the verification result is fed back to the visual feature adapter to adjust the visual features and continue to perform task selection and performance verification. If the determination result is positive, the training of the thought chain model ends.

[0049] The visual feature adapter is used to convert the general visual features output by the visual-language large model into a specific format required by the thought chain inference model, and to dynamically adjust the visual features based on the feedback results of the thought chain performance verification.

[0050] The functions of the thought chain reasoning model include: decomposing the task according to the task description language, determining the reasoning steps, and selecting the optimal downstream visual network for each reasoning operation.

[0051] Downstream visual networks include at least one or more of the following networks: change detection network, change description network, object recognition network, and semantic segmentation network.

[0052] In operation S240, change detection processing is performed on multi-temporal remote sensing images based on the reasoning results of the thought chain task, and the change detection results of multi-temporal remote sensing images are obtained.

[0053] The multi-temporal remote sensing images to be tested and the change understanding task description are sequentially processed through language encoding, a large visual language model, a visual feature adapter network, and a thought chain reasoning model. At the end of the iteration of the thought chain reasoning model, the results of each subtask constitute the final change detection and reasoning understanding results.

[0054] The aforementioned change detection results are structured data, namely, the results of understanding the changes in multi-temporal remote sensing images at different times, including the location information of the changed areas, the natural language description of the changes, and the object categories and attribute information related to the changes.

[0055] The remote sensing image change detection method provided by this invention aligns textual information with visual features through semantic vectors encoded by a task description language, effectively solving the false detection problem caused by radiometric differences and registration errors in traditional methods. It achieves multimodal collaborative enhancement in the change detection and inference process of multi-temporal remote sensing images. Simultaneously, by utilizing a trained thought chain model through step-by-step inference, it can distinguish between real changes and interference factors in complex scenes and dynamically allocate computing resources according to task requirements, improving processing efficiency and reducing computational costs. Furthermore, the remote sensing image change detection method provided by this invention displays the collaborative inference process of multiple tasks through the thought chain, meeting users' needs for interpretability, greatly improving the effect and efficiency of remote sensing image processing, and providing a more convenient and reliable service for practical applications.

[0056] According to an embodiment of the present invention, the acquisition of multi-temporal remote sensing images of the same imaging target includes: acquiring a first remote sensing image of the same imaging target at a first time and a second remote sensing image at a second time; performing radiometric correction and registration processing on the first remote sensing image to obtain a preprocessed first remote sensing image, and performing radiometric correction and registration processing on the second remote sensing image to obtain a preprocessed second remote sensing image.

[0057] Remote sensing images of the same area at time t1 and time t2 are acquired to obtain multi-temporal remote sensing images of the same area; the multi-temporal remote sensing images are preprocessed to perform radiometric correction and registration; at the same time, a task description language is set to describe the analysis requirements, such as "detect changes in buildings".

[0058] According to an embodiment of the present invention, the above-mentioned language encoding processing of the task description language to obtain semantic vectors includes: performing word vector embedding fusion operation, bidirectional Transformer encoding operation, and pooling operation on words related to land features in the task description language to obtain semantic vectors; wherein, based on the semantic vectors, feature extraction of multi-temporal remote sensing images to obtain unified visual features includes: using a pre-trained visual-language large model to perform cross-modal feature alignment operation based on contrastive learning and mapping operation based on the same semantic space on the first remote sensing image and the second remote sensing image to obtain unified visual features, wherein the same semantic space is represented by the semantic vectors.

[0059] Terminology related to geographical features generally refers to relatively fixed objects on the Earth's surface, including man-made or naturally formed objects, such as a building in a remote sensing image. Encoding the task description language essentially involves encoding terminology related to geographical features within the task description language. For example, "detecting changes in buildings" actually encodes the word "building," which is a term related to geographical features. Terminology related to geographical features changes over time in remote sensing images and thus belongs to the category of change-type terms.

[0060] The process of obtaining uniform visual features will be further explained in detail below through specific implementation methods.

[0061] Unified visual feature extraction: The remote sensing images at time t1 and time t2, along with the language-encoded task description, are input into a pre-trained visual-language large model to extract unified visual features.

[0062] First, the task description language is encoded: the task description language... (For example, "detecting building changes") generates semantic vectors using the language encoding technique BERT (Bidirectional Encoder Representations from Transformers). ,Right now .in, For task description language, This is the semantic vector generated by the BERT encoder. Its function is to provide language cues for visual language models and task cues for the decomposition of thought chain tasks.

[0063] Secondly, a unified visual feature extraction method is used: Change understanding is based on reasoning and collaboration among sub-tasks such as change detection, change description, semantic segmentation, and object detection using visual language features. Considering the similarity among the sub-tasks of change understanding, they share a common visual feature backbone network, and the pre-training of these sub-tasks only trains the head network parameters of the sub-tasks. Specifically, for dual-temporal remote sensing images... , Based on semantic vectors Extracting visual features from the CLIP model .

[0064] The above embodiments employ bidirectional Transformer encoding of the task description language, capturing long-range dependencies between words through a self-attention mechanism to address the problem of polysemy that traditional word vectors cannot handle. The combination of word vector embedding fusion and pooling operations preserves local semantic details (such as technical terms) while integrating task-level intents through global pooling, forming a hierarchical semantic representation. Simultaneously, the visual-language large model maps image segmentation results and text feature vectors to the same semantic space, achieving intermodal feature alignment through contrastive learning. The multi-scale processing capability of the pre-trained model adapts to the large scale differences of ground features in remote sensing images, simultaneously capturing local texture (through CNN) and global context (through Transformer).

[0065] According to an embodiment of the present invention, the trained visual feature adapter and the trained thought chain inference model are obtained through the following operations: using the visual feature adapter to perform super-resolution image reconstruction on the visual feature samples of multi-temporal remote sensing image samples to obtain preprocessed visual feature samples; pre-training the visual feature adapter and multiple sub-task models for multi-temporal remote sensing image processing to obtain a pre-trained visual feature adapter and multiple pre-trained sub-task models; iteratively training the thought chain inference model and the pre-trained visual feature adapter using semantic vector samples corresponding to the visual feature samples, the preprocessed visual feature samples, and the multiple pre-trained sub-task models to obtain a trained visual feature adapter and a trained thought chain inference model.

[0066] In the pre-training phase of the above embodiments, in addition to super-resolution reconstruction, more self-supervised sub-tasks can be introduced, such as image colorization, rotation prediction, and season invariance learning. This forces the model to learn more robust and general visual representations, rather than just pixel-level details. The Chain of Thought (CoT) model is upgraded from linear inference steps to a dynamic task inference graph. The CoT inference model can dynamically select and combine pre-trained sub-task models (e.g., "building segmentation" → "variable region extraction" → "variable type classification") based on the input semantic vector and visual features, forming an optimal inference path for a specific task.

[0067] According to an embodiment of the present invention, the above-mentioned pre-training of the visual feature adapter and multiple sub-task models for multi-temporal remote sensing image processing to obtain a pre-trained visual feature adapter and multiple pre-trained sub-task models includes: using the visual feature adapter with randomly initialized parameters to perform channel-dimensional concatenation processing on the multi-temporal visual features of the multi-temporal remote sensing image samples to obtain multi-temporal feature map samples; processing the multi-temporal feature map samples using multiple sub-task models to obtain predicted values ​​representing different processing tasks of the multi-temporal remote sensing image samples; calculating the loss between the predicted values ​​and the true values ​​of the multi-temporal remote sensing image samples to obtain pre-trained loss values; and updating the parameters of the visual feature adapter and multiple sub-task models using the pre-trained loss values ​​through backpropagation to obtain the pre-trained visual feature adapter and multiple pre-trained sub-task models.

[0068] In the initial pre-training phase of the above embodiments, different weights can be assigned to different sub-tasks, prioritizing training tasks more closely related to the core objective (such as change detection) (e.g., semantic segmentation). Once the model stabilizes, other auxiliary tasks (such as super-resolution reconstruction and cloud detection) can be gradually introduced to stabilize the training process and improve final performance. Furthermore, based on the multi-temporal feature maps obtained through channel concatenation, the above embodiments can introduce cross-temporal comparative learning. Moreover, the joint pre-training mechanism employed in the above embodiments constructs a powerful, general-purpose, and efficient basic model for remote sensing image analysis, laying a solid foundation for subsequent thought chain reasoning and accurate change detection.

[0069] According to embodiments of the present invention, the above-mentioned sub-task models include a change detection task model, a change description task model, a target recognition task model, and a semantic segmentation task model.

[0070] According to an embodiment of the present invention, the above-mentioned iterative training of the thought chain inference model and the pre-trained visual feature adapter using semantic vector samples corresponding to visual feature samples, pre-processed visual feature samples, and multiple pre-trained sub-task models to obtain the trained visual feature adapter and the trained thought chain inference model includes: in the first iteration, based on the semantic vector samples, the thought chain model is used to perform task decomposition, and based on the result of task decomposition, all sub-task models are called to perform processing on multi-temporal remote sensing image samples; in other iterations, based on the semantic vector samples, the pre-processed visual feature samples and multiple pre-trained sub-task models are used to iteratively perform task selection and performance verification operations based on reinforcement learning algorithms on the thought chain model and the pre-trained visual feature adaptation to obtain the trained visual feature adapter and the trained thought chain inference model.

[0071] According to an embodiment of the present invention, in the above-mentioned other iteration processes, based on semantic vector samples, using preprocessed visual feature samples and multiple pre-trained sub-task models, iteratively performing task selection and performance verification operations based on reinforcement learning algorithms on the mental chain model and the pre-trained visual feature adaptation to obtain a trained visual feature adapter and a trained mental chain inference model includes: based on semantic vector samples, performing task inference operations on preprocessed visual feature samples using the mental chain model to obtain task inference results; selecting a pre-trained target task model from multiple pre-trained sub-task models based on the task inference results, and using the pre-trained target task model to perform image processing on multi-temporal remote sensing image samples to obtain intermediate results; performing verification operations on the intermediate processing results, and performing parameter update operations on the pre-trained visual feature adapter and mental chain inference model based on the verification results to obtain a trained visual feature adapter and a trained mental chain inference model.

[0072] The above embodiments achieve collaborative optimization of the visual feature adapter and the thought chain inference model through a reinforcement learning-driven task selection and performance verification mechanism: the thought chain model dynamically selects the most relevant sub-task model based on semantic vectors and current visual features (e.g., selecting the "building segmentation" task instead of the "road extraction" task), avoiding redundant calculations for all sub-tasks and significantly improving inference efficiency; reinforcement learning reward signals are generated by verifying intermediate results (e.g., segmentation accuracy, change detection consistency), directly guiding the parameter updates of the visual feature adapter and the thought chain model. This closed-loop design of "inference-verification-feedback" enables the model to quickly converge to the optimal solution; the task inference results output by the thought chain model (e.g., "detecting new buildings") are semantically aligned with the features extracted by the visual feature adapter, ensuring that the sub-task model processes image regions highly relevant to the text description, reducing false positives and false negatives.

[0073] According to an embodiment of the present invention, the above-mentioned selection of a pre-trained target task model from multiple pre-trained sub-task models based on task inference results includes: determining the current state in the current training round based on task inference results and the output results of the sub-task models in the previous training round; determining the selection action based on the current state and the value function; and selecting the pre-trained target task model from multiple pre-trained sub-task models based on the selection action.

[0074] The training process of the above-mentioned thinking chain model provided by the present invention will be further explained in detail below through specific implementation methods.

[0075] Unified visual features are input into the thought chain inference model via a visual feature adapter. The thought chain inference model decomposes the image change understanding task into one or more sub-tasks and selectively calls at least one preset downstream visual network to execute the corresponding sub-task, generating intermediate results. The intermediate results are then subjected to performance verification to determine whether the inference process meets preset termination conditions. If the determination result is negative, the verification result is fed back to the visual feature adapter to adjust the visual features and continue task selection and performance verification. If the determination result is positive, the thought chain model training ends. The above operations further include the following:

[0076] First, regarding the visual feature adapter: visual features The feature map for each channel is smaller than the original multi-temporal image, which is inconsistent with the feature size of the change mask and semantic segmentation mask. This makes it impossible to directly train the sub-task network, and the detailed information of the changes is difficult to learn. The visual feature adapter aims to learn the changes through a simple three-layer fully connected network. Perform super-resolution reconstruction to achieve a feature map of the same size as the change mask and semantic segmentation mask. The network parameters of the visual feature adapter are learned together with the network model pre-training operations of the following subtasks.

[0077] Secondly, pre-training of the change comprehension sub-task model: To facilitate the training of the thought chain model, the four sub-task network models are pre-trained according to their corresponding annotations. Specifically, the feature adapter network parameters are first randomly initialized to obtain... ; As input to semantic segmentation and object detection networks, multi-temporal semantic segmentation and object detection results are obtained. By comparing semantic segmentation maps at different times, semantic changes can be inferred; by comparing object detection results at different times, object changes can be inferred. Multi-temporal visual features are obtained by connecting the channels. Multi-temporal visual features Inputting the change detection and change description subtask networks yields a change mask and a textual description of the change. Based on the combined loss of the predicted and true values ​​from each subtask network, the backpropagation algorithm is used to iteratively refine the parameters of the feature adapter network and each subtask network, resulting in a pre-trained model for each subtask.

[0078] Finally, the thought chain model is trained: the thought chain reasoning model is a unified visual feature extraction operation definition that understands the switching and mutual verification of subtasks. In the thought chain reasoning model of this invention, the switching and verification of subtasks are implemented by reinforcement learning. The thought chain model training process includes the following operations:

[0079] First, task decomposition: Task decomposition is used during the first iteration of the training process of the thought chain reasoning model. In the first iteration, all four sub-tasks need to be executed. The execution instructions for the sub-tasks are defined by the task semantic vectors encoded by the thought chain reasoning model and the task description language. To complete it together.

[0080] Secondly, task selection and performance verification: Task selection is used when the number of iterations exceeds 2 during the training of the thought chain inference model. Task selection chooses only one of the four subtasks. This invention uses the reinforcement learning algorithm DQN (Deep Q-Network) to dynamically select and switch subtasks. Specifically, the state... This is the output and context information of the current subtask. Action Select the next subtask. Value function Indicates the state Take action below The expected return. Defined Understanding the instantaneous rewards of subtask consistency in light of changing moments: ,in, For subtasks The weight, For the first The performance metrics for each task are as follows: change detection, semantic segmentation, and object detection are measured by mIoU (an evaluation metric in semantic segmentation), while change description is measured by BLEU-4 (an evaluation metric in natural language processing). The KL divergence represents the difference between the current task result and the previous task result. The value function is updated as shown in Equation (1):

[0081] (1).

[0082] in, It's the learning rate. It is a discount factor. The next state is... This is the next action.

[0083] The specific training process for task selection is as follows:

[0084] initialization Value function ;

[0085] For each pair of remote sensing images and Repeat the following steps until the termination condition is met or the predetermined training cycle is completed:

[0086] Operation 1: Select initial state ;

[0087] Operation 2, for each time step :

[0088] Operation 2-1, based on the current state and Value function selection action ;

[0089] Operation 2-2, perform the action Observe the next state and rewards ;

[0090] Operations 2-3: Update the Q-value function;

[0091] Operations 2-4: Verify the performance of the thought chain. If the performance of each subtask in the current change understanding is less than the given threshold, the task selection is terminated, and the output of each subtask is output; otherwise, proceed to step 2-5.

[0092] Operation 2-5: Based on the difference between the prediction result and the actual result of the current subtask, update the parameters of the visual feature adapter network and execute operation 2-1.

[0093] In reinforcement learning, the reward function incorporates changes in understanding the performance of subtasks and the consistency between them. The consistency between subtasks is represented by the KL divergence between the results of the current task and the results of the previous task.

[0094] Figure 3 A training and testing framework diagram for remote sensing image change detection and understanding based on thought chain reasoning according to an embodiment of the present invention is shown.

[0095] like Figure 3As shown, images at time t1 (i.e., remote sensing images at time t1) and time t2 (i.e., remote sensing images at time t2) are acquired, along with a task description language describing the analysis requirements. The images and the encoded task description language are input into a large-scale vision-language model to extract unified visual features. These visual features are then input into a thought chain inference model via a visual feature adapter. This model decomposes the complex transformation understanding task into a series of sub-tasks and selectively calls one or more downstream visual networks (such as change detection, change description, object recognition, and semantic segmentation networks) to perform these tasks. The output results of the downstream networks are verified using thought chain performance to determine if the inference process has ended. If not, the verification results are fed back to the visual feature adapter for iterative optimization until the inference process ends. Finally, the structured transformation understanding results are integrated and output. This invention significantly improves the depth, accuracy, and logic of understanding image changes in complex scenes by simulating the human thought chain process, combining the general capabilities of a large-scale model with the accuracy of a specialized model, and using an iterative verification mechanism for self-correction. During training, the visual feature adapter and the reasoning performance verification of the thought chain form a closed-loop feedback loop for iterative optimization and self-correction of the reasoning process. The downstream visual network includes at least one or more of the following networks: change detection network, change description network, object recognition network, and semantic segmentation network. Each subtask of change understanding shares a visual feature backbone network, and the pre-training of the change understanding subtasks only trains the parameters of the subtask's head network.

[0096] Based on the above-described remote sensing image change detection method, this invention also provides a change detection device. The following will be combined with... Figure 4 The device is described in detail.

[0097] Figure 4 A structural block diagram of a remote sensing image change detection device according to an embodiment of the present invention is shown.

[0098] like Figure 4 As shown, the remote sensing image change detection device 400 of this embodiment includes an image acquisition and task setting module 410, a visual feature acquisition module 420, a thought chain reasoning module 430, and a multi-temporal remote sensing image processing module 440.

[0099] The image acquisition and task setting module 410 is used to acquire multi-temporal remote sensing images of the same imaging target and set a task description language to characterize the multi-modal processing of multi-temporal remote sensing images. In one embodiment, the image acquisition and task setting module 410 can be used to perform the operation S210 described above, which will not be repeated here.

[0100] The visual feature acquisition module 420 is used to perform language encoding processing on the task description language to obtain semantic vectors, and to extract features from multi-temporal remote sensing images based on the semantic vectors to obtain unified visual features. In one embodiment, the visual feature acquisition module 420 can be used to perform the operation S220 described above, which will not be repeated here.

[0101] The thought chain reasoning module 430 is used to preprocess the unified visual features using the trained visual feature adapter to obtain the preprocessed visual features, and to perform task reasoning on the preprocessed visual features based on the semantic vector using the trained thought chain model to obtain the thought chain task reasoning result. In one embodiment, the thought chain reasoning module 430 can be used to execute the operation S230 described above, which will not be repeated here.

[0102] The multi-temporal remote sensing image processing module 440 is used to perform change detection processing on the multi-temporal remote sensing images based on the reasoning results of the thought chain task, and obtain the change detection results of the multi-temporal remote sensing images. In one embodiment, the multi-temporal remote sensing image processing module 440 can be used to perform the operation S240 described above, which will not be repeated here.

[0103] According to embodiments of the present invention, any multiple modules of the image acquisition and task setting module 410, visual feature acquisition module 420, thought chain reasoning module 430, and multi-temporal remote sensing image processing module 440 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of the present invention, at least one of the image acquisition and task setting module 410, visual feature acquisition module 420, thought chain reasoning module 430, and multi-temporal remote sensing image processing module 440 can be at least partially implemented as hardware circuitry, such as field-programmable gate array (FPGA), programmable logic array (PLA), system-on-a-chip, system-on-a-substrate, system-on-package, application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the image acquisition and task setting module 410, visual feature acquisition module 420, thought chain reasoning module 430, and multi-temporal remote sensing image processing module 440 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0104] Figure 5 A block diagram of an electronic device suitable for implementing a remote sensing image change detection method according to an embodiment of the present invention is shown.

[0105] like Figure 5 As shown, an electronic device 500 according to an embodiment of the present invention includes a processor 501, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 502 or a program loaded from a storage portion 508 into a random access memory (RAM) 503. The processor 501 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 501 may also include onboard memory for caching purposes. The processor 501 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0106] RAM 503 stores various programs and data required for the operation of electronic device 500. Processor 501, ROM 502, and RAM 503 are interconnected via bus 504. Processor 501 executes various operations of the method flow according to embodiments of the present invention by executing programs in ROM 502 and / or RAM 503. It should be noted that the programs may also be stored in one or more memories other than ROM 502 and RAM 503. Processor 501 may also execute various operations of the method flow according to embodiments of the present invention by executing programs stored in said one or more memories.

[0107] According to an embodiment of the present invention, the electronic device 500 may further include an input / output (I / O) interface 505, which is also connected to a bus 504. The electronic device 500 may also include one or more of the following components connected to the input / output (I / O) interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN card, modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to the input / output (I / O) interface 505 as needed. A removable medium 511, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 510 as needed so that computer programs read from it can be installed into the storage section 508 as needed.

[0108] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present invention.

[0109] According to embodiments of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of the present invention, the computer-readable storage medium may include ROM 502 and / or RAM 503 and / or one or more memories other than ROM 502 and RAM 503 described above.

[0110] Embodiments of the present invention also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code enables the computer system to implement the remote sensing image change detection method provided in the embodiments of the present invention.

[0111] When the computer program is executed by the processor 501, it performs the functions defined in the system / apparatus of this invention. According to embodiments of the invention, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0112] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 509, and / or installed from a removable medium 511. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0113] In such an embodiment, the computer program can be downloaded and installed from a network via communication section 509, and / or installed from removable medium 511. When the computer program is executed by processor 501, it performs the functions defined in the system of this embodiment of the invention. According to embodiments of the invention, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0114] According to embodiments of the present invention, program code for executing the computer programs provided in the embodiments of the present invention can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0115] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0116] Those skilled in the art will understand that the features described in the various embodiments of the present invention can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention can be combined and / or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

[0117] The embodiments of the present invention have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the invention, and all such substitutions and modifications should fall within the scope of the invention.

Claims

1. A method for detecting changes in remote sensing images, characterized in that, The method includes: Acquire multi-temporal remote sensing images of the same imaging target, and set a task description language to characterize the multi-modal processing required for the multi-temporal remote sensing images; The task description language is coded to obtain a semantic vector, and features are extracted from the multi-temporal remote sensing images based on the semantic vector to obtain unified visual features. The unified visual features are preprocessed using the trained visual feature adapter to obtain preprocessed visual features. Based on the semantic vector, the trained thought chain model is used to perform task reasoning on the preprocessed visual features to obtain the thought chain task reasoning result. Based on the reasoning results of the thought chain task, change detection processing is performed on the multi-temporal remote sensing images to obtain the change detection results of the multi-temporal remote sensing images.

2. The method according to claim 1, characterized in that, Acquiring multi-temporal remote sensing images of the same imaging target includes: Acquire a first remote sensing image of the same imaging target at a first time and a second remote sensing image at a second time, respectively; The first remote sensing image is subjected to radiometric correction and registration processing to obtain a preprocessed first remote sensing image, and the second remote sensing image is subjected to radiometric correction and registration processing to obtain a preprocessed second remote sensing image.

3. The method according to claim 2, characterized in that, The task description language is subjected to language encoding processing to obtain a semantic vector, including: Perform word vector embedding and fusion, bidirectional Transformer encoding, and pooling operations on the words related to ground features in the task description language to obtain semantic vectors; Specifically, based on the semantic vector, feature extraction is performed on the multi-temporal remote sensing images to obtain unified visual features, including: The pre-trained visual-language large model is used to perform cross-modal feature alignment based on contrastive learning and mapping operation based on the same semantic space on the first remote sensing image and the second remote sensing image to obtain unified visual features, wherein the same semantic space is represented by the semantic vector.

4. The method according to claim 1, characterized in that, The trained visual feature adapter and the trained thought chain reasoning model are obtained through the following operations: Super-resolution image reconstruction was performed on the visual feature samples of multi-temporal remote sensing image samples using a visual feature adapter to obtain preprocessed visual feature samples. The visual feature adapter and multiple sub-task models for multi-temporal remote sensing image processing are pre-trained to obtain a pre-trained visual feature adapter and multiple pre-trained sub-task models. The thought chain reasoning model and the pre-trained visual feature adapter are iteratively trained using semantic vector samples corresponding to the visual feature samples, the pre-processed visual feature samples, and multiple pre-trained sub-task models to obtain the trained visual feature adapter and the trained thought chain reasoning model.

5. The method according to claim 4, characterized in that, The visual feature adapter and multiple sub-task models for multi-temporal remote sensing image processing are pre-trained to obtain a pre-trained visual feature adapter and multiple pre-trained sub-task models, including: The multi-temporal visual features of the multi-temporal remote sensing image samples are connected along the channel dimension using a visual feature adapter with randomly initialized parameters to obtain multi-temporal feature map samples. The multi-temporal feature map samples are processed using multiple sub-task models to obtain predicted values ​​that characterize the multi-temporal remote sensing image samples for different processing tasks. The predicted value is compared with the true value of the multi-temporal remote sensing image sample to calculate the loss, and the pre-training loss value is obtained. The pre-trained loss value is used to update the parameters of the visual feature adapter and the multiple sub-task models through backpropagation to obtain the pre-trained visual feature adapter and the multiple pre-trained sub-task models.

6. The method according to claim 5, characterized in that, The sub-task models include a change detection task model, a change description task model, an object recognition task model, and a semantic segmentation task model.

7. The method according to claim 4, characterized in that, The thought chain reasoning model and the pre-trained visual feature adapter are iteratively trained using semantic vector samples corresponding to the visual feature samples, the preprocessed visual feature samples, and multiple pre-trained sub-task models to obtain the trained visual feature adapter and the trained thought chain reasoning model, including: In the first iteration, based on the semantic vector samples, the thought chain model is used to decompose the task, and based on the result of the task decomposition, all the sub-task models are called to process the multi-temporal remote sensing image samples. In other iterations, based on the semantic vector samples, the preprocessed visual feature samples and multiple pre-trained sub-task models are used to iteratively perform task selection and performance verification operations based on reinforcement learning algorithms on the mind chain model and the pre-trained visual feature adaptability, so as to obtain the trained visual feature adapter and the trained mind chain reasoning model.

8. The method according to claim 7, characterized in that, In other iterations, based on the semantic vector samples, the preprocessed visual feature samples and multiple pre-trained sub-task models are used to iteratively perform task selection and performance verification operations based on reinforcement learning algorithms on the mind chain model and the pre-trained visual feature adaptation, resulting in the trained visual feature adapter and the trained mind chain reasoning model, including: Based on the semantic vector samples, the thought chain model is used to perform task reasoning operations on the preprocessed visual feature samples to obtain task reasoning results. Based on the task inference results, a pre-trained target task model is selected from multiple pre-trained sub-task models, and the pre-trained target task model is used to perform image processing on the multi-temporal remote sensing image samples to obtain intermediate results. A verification operation is performed on the intermediate processing results, and a parameter update operation is performed on the pre-trained visual feature adapter and the thought chain reasoning model based on the verification results, so as to obtain the trained visual feature adapter and the trained thought chain reasoning model.

9. The method according to claim 8, characterized in that, Selecting a pre-trained target task model from multiple pre-trained sub-task models based on the task inference results includes: The current state in the current training round is determined based on the task inference results and the output results of the sub-task model in the previous training round. The selection action is determined based on the current state and value function, and the pre-trained target task model is selected from multiple pre-trained sub-task models based on the selection action.

10. A remote sensing image change detection device, characterized in that, The device includes: The image acquisition and task setting module is used to acquire multi-temporal remote sensing images of the same imaging target and set a task description language to characterize the multi-modal processing required for the multi-temporal remote sensing images. The visual feature acquisition module is used to perform language encoding processing on the task description language to obtain a semantic vector, and to extract features from the multi-temporal remote sensing images based on the semantic vector to obtain unified visual features. The thinking chain reasoning module is used to preprocess the unified visual features using the trained visual feature adapter to obtain preprocessed visual features, and to perform task reasoning on the preprocessed visual features using the trained thinking chain model based on the semantic vector to obtain the thinking chain task reasoning result. The multi-temporal remote sensing image processing module is used to perform change detection processing on the multi-temporal remote sensing image based on the reasoning result of the thought chain task, and obtain the change detection result of the multi-temporal remote sensing image.

Images