Remote sensing image building damage evaluation criterion optimization method and system

Abstract

The invention discloses a remote sensing image building damage assessment criterion optimization method, which comprises the steps of obtaining a typical target damage case library and an initial damage assessment criterion, the typical target damage case library comprising qualitative damage features of a building and artificially calibrated building damage levels; inputting the typical target damage case library and the initial damage evaluation criterion into the language large model to obtain a corresponding target damage level; calculating damage precision and screening damage cases based on the target damage grade and the artificially calibrated building damage grade, and generating a new damage evaluation criterion by using the screened new damage cases; and according to the new damage assessment criterion and the new damage case, re-optimizing the damage assessment criterion until the damage precision does not increase any more, and determining the latest damage assessment criterion as the optimized damage assessment criterion. On the basis, continuous iteration and dynamic optimization of the damage evaluation criterion are realized.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and computer natural language processing, specifically to a method and system for optimizing building damage assessment criteria based on remote sensing images. Background Technology

[0002] In recent years, the rapid development of science and technology has greatly promoted the diversification and intelligence of target engagement equipment in the military field. From precision-guided weapons to unmanned combat platforms, from hypersonic missiles to artificial intelligence-based command and control systems, the enrichment of strike methods has not only significantly improved combat effectiveness but also made the damage patterns of targets after being attacked more complex and varied. This diversity of damage effects is specifically reflected in the coexistence of multiple forms such as structural collapse, combustion and explosion, partial disintegration, deformation and tilting, and even the emergence of multi-level and multi-type composite damage modes, posing unprecedented challenges to accurately assessing the damage level of targets.

[0003] However, the currently widely adopted damage assessment criteria system still relies heavily on traditional experience and judgment rules under limited scenarios, exhibiting significant deficiencies in both the breadth and depth of knowledge coverage. Many novel damage modes have not yet been systematically incorporated into the assessment criteria. This lag in the knowledge system leads to assessment results that often deviate from the actual damage state in practical applications, resulting in a substantial decrease in accuracy. Inaccurate assessment results not only directly affect the cost-effectiveness of a single strike operation but also severely constrain subsequent operational decisions. Command organizations are heavily reliant on reliable assessments of the initial strike's effectiveness when deciding whether to launch a second strike, how to select aiming points, and how to rationally allocate ammunition and combat platforms. The current imperfections in the assessment criteria significantly increase decision-making uncertainty, potentially leading to excessive strikes resulting in resource waste or insufficient firepower failing to achieve strategic objectives, thereby impacting the entire operational process. Summary of the Invention

[0004] To overcome the shortcomings of existing building damage assessment criteria that still rely on traditional experience and judgment rules under limited scenarios, this invention provides a method and system for optimizing building damage assessment criteria based on remote sensing images. It integrates advanced computing models with remote sensing image analysis capabilities, and continuously improves the judgment accuracy and robustness of the assessment system by continuously iterating and dynamically optimizing the damage assessment criteria.

[0005] According to one aspect of the present invention, a method for optimizing building damage assessment criteria in remote sensing imagery is provided, comprising: step S1, acquiring a typical target damage case library and initial damage assessment criteria, wherein the typical target damage case library includes qualitative damage characteristics of buildings and manually labeled building damage levels; step S2, inputting the typical target damage case library and initial damage assessment criteria into a language model to obtain corresponding target damage levels; wherein the initial damage assessment criteria are incorporated into the language model in the form of prior knowledge, so that the language model determines the damage level of buildings based on the qualitative damage characteristics of the buildings according to the initial damage assessment criteria; step S3, calculating damage accuracy and screening damage cases based on the target damage levels and manually labeled building damage levels, and generating new damage assessment criteria using the screened new damage cases; step S4, re-executing steps S2 to S3 based on the new damage assessment criteria and new damage cases until the damage accuracy no longer increases, and recognizing the latest damage assessment criteria as the optimized damage assessment criteria.

[0006] Furthermore, the initial damage assessment criteria are incorporated into the language big model in the form of prior knowledge, including: inputting the initial damage assessment criteria into the trained embedded model to obtain the corresponding text representation vector; inputting the text representation vector as prior knowledge into the language big model and storing it in the local database of the language big model.

[0007] Furthermore, new damage assessment criteria are generated using the selected new damage cases, including: extracting features and performing semantic analysis on the new damage cases through a large language model to generate high-dimensional vector representations for association and comparison with the initial damage assessment criteria; after completing the association and comparison with the initial damage assessment criteria, the new damage cases are summarized using a large language model to form new damage assessment criteria.

[0008] Furthermore, after obtaining the optimized damage assessment criteria, the optimization method further includes: inputting remote sensing images before and after damage into a building damage semantic description model to obtain target damage features; and inputting the target damage features and the optimized damage assessment criteria into a large language model to obtain the target damage level.

[0009] Furthermore, the semantic description model of building damage includes an image and text information extraction module, a feature information fusion module, and a text generation module; the image and text information extraction module is used to extract key feature information from remote sensing images before and after damage, the key feature information including image feature information and text feature information; the feature information fusion module is used to fuse image feature information and text feature information at multiple scales to obtain fused feature information; the text generation module is used to generate corresponding text based on the fused feature information.

[0010] Furthermore, the image and text information extraction module includes a large remote sensing image and text alignment model and a LoRA layer, the formula corresponding to the LoRA layer being:

[0011] ,

[0012] in, The input features of the LoRA layer are represented by A and B, which are two updatable weight parameters, and W represents the pre-trained weights. This represents the weight increment that needs fine-tuning. The dot product of B and A. It is a scaling factor. Represents dot product. Represents matrix multiplication. This represents the input features of the LoRA layer.

[0013] Furthermore, the feature information fusion module includes an image-text fusion part and a spatiotemporal fusion part. In the image-text fusion part, the Transformer module and multimodal feature retrieval are used to complete the spatial alignment and semantic information alignment of cross-modal features, and a unified semantic space is constructed using learnable affine transformations. In the spatiotemporal fusion part, a multi-head cross-attention mechanism is used to establish fine-grained associations between modalities, and dynamic information interaction is achieved by combining gated feature filtering and residual fusion. The Mamba attention mechanism is used to capture long-range dependencies, and the relationship between multi-temporal features and the correspondence rules between semantic information and text information are established through feature weighting and residual connections.

[0014] Furthermore, the text generation module includes a masked multi-head attention mechanism and a Transformer decoder. The formulas corresponding to the Transformer module and the Mamba attention mechanism are as follows:

[0015] ,

[0016] ,

[0017] in,( ) represents the input features of the Transformer module. for Input features of the attention mechanism; Represents a feedforward neural network. Represents the activation function. Represents a fully connected layer. Represents dot product. Represents matrix multiplication. It is a constant. Represents the state space. This represents a convolutional layer.

[0018] According to one aspect of the present invention, a remote sensing image building damage assessment criterion optimization system is provided, comprising: a data acquisition module for acquiring a typical target damage case library and initial damage assessment criteria, wherein the typical target damage case library includes qualitative damage characteristics of buildings and manually labeled building damage levels; a damage level determination module for inputting the typical target damage case library and initial damage assessment criteria into a language large model to obtain corresponding target damage levels; wherein the initial damage assessment criteria are incorporated into the language large model in the form of prior knowledge, so that the language large model determines the damage level of buildings based on the qualitative damage characteristics of the buildings according to the initial damage assessment criteria; an accuracy calculation and case screening module for calculating damage accuracy and screening damage cases based on the target damage levels and manually labeled building damage levels, and generating new damage assessment criteria using the screened new damage cases; and a damage assessment criterion optimization module for re-optimizing the damage assessment criteria based on the new damage cases and the new damage criteria until the damage accuracy no longer increases, and recognizing the latest damage assessment criteria as the optimized damage assessment criteria.

[0019] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the method for optimizing the assessment criteria for building damage from remote sensing images.

[0020] The aforementioned technical solution comprises two parts: training and testing. The training phase mainly includes: creating a database of building damage cases; determining the damage level of buildings based on qualitative damage characteristics according to damage assessment criteria; screening damage cases; summarizing and generalizing using a large language model to generate new damage assessment criteria; and iteratively optimizing the results using inference accuracy as supervision to obtain the final optimized building damage assessment criteria. The testing phase mainly includes: semantic description of building damage; determining the damage level of buildings based on qualitative damage characteristics according to the optimized damage assessment criteria, achieving accurate determination of building damage levels. This invention embodies a collaborative mechanism of "data-damage assessment criteria-model," emphasizing both the objectivity of remote sensing imagery and the guiding role of domain knowledge, while leveraging a large language model to improve the analysis efficiency and generalization ability of complex scenarios.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] (1) The scope of damage assessment criteria has been expanded. With the powerful reasoning and inductive capabilities of the large language model, the system can extract damage features other than typical structural damage (such as craters and collapses) from a large amount of expert knowledge and actual cases. These features include, but are not limited to, the target's burning state, degree of fragmentation and disintegration, and subtle morphological changes such as structural deformation and tilt. Through the standardized description of these explicit damage indicators, the completeness and expressiveness of the damage assessment criteria are significantly improved.

[0023] (2) Self-iteration and optimization of damage assessment criteria are achieved. This invention performs preliminary damage assessment on remote sensing images based on existing damage assessment criteria, and compares the assessment results with cases already labeled by experts, selecting correctly assessed cases as positive samples. On this basis, preliminary optimized damage assessment criteria are generated through logical reasoning and rule synthesis. This process is executed cyclically, continuously introducing new expert samples and assessment feedback, and approaching the optimal solution for assessment accuracy through multiple rounds of iteration, ultimately forming a set of assessment criteria systems highly adapted to actual combat scenarios. This method not only improves the accuracy and timeliness of damage assessment, but also enhances the adaptability of the assessment system in complex and ever-changing battlefield environments, providing more reliable intelligence support for command and decision-making. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating an optimization method for assessing building damage criteria using remote sensing imagery, provided as an embodiment of the present invention.

[0026] Figure 2 A flowchart illustrating the building damage level inference technology provided in this embodiment of the invention.

[0027] Figure 3 A flowchart illustrating the optimization technology for building damage assessment criteria provided in this embodiment of the invention.

[0028] Figure 4 The diagram shows the structure of the KGBDNet model provided in this embodiment of the invention.

[0029] Figure 5 A comparison diagram showing the optimized building damage criteria provided in this embodiment of the invention.

[0030] Figure 6 The results of the reasoning accuracy of the damage criteria before and after optimization are provided for the embodiments of the present invention. Detailed Implementation

[0031] It should be noted that:

[0032] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0033] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0035] Please refer to the appendix. Figure 1 This invention provides a method for optimizing building damage assessment criteria based on remote sensing imagery. This method consists of two phases: a training phase and a testing phase. The training phase optimizes the damage assessment criteria for typical targets, while the testing phase assesses the damage level of typical targets. In the training phase, a database of typical target damage cases and initial damage assessment criteria are used as inputs, and the optimized damage assessment criteria are output. The training phase includes the following steps:

[0036] Step S1: Obtain a typical target damage case library and initial damage assessment criteria. The typical target damage case library includes the qualitative damage characteristics of buildings and manually calibrated building damage levels.

[0037] Step S2: Input the typical target damage case library and the initial damage assessment criteria into the language big model to obtain the corresponding target damage level; wherein, the initial damage assessment criteria are incorporated into the language big model in the form of prior knowledge, so that the language big model can determine the damage level of the building based on the qualitative damage characteristics of the building according to the initial damage assessment criteria.

[0038] In this embodiment, the pre-trained large model Qwen3 is loaded as the language large model of this invention. It should be noted that typical target damage level determination techniques involve complex physical characteristic analysis and professional domain knowledge. Utilizing a language large model in conjunction with target damage assessment criteria is key to achieving high-precision and high-reliability damage inference. The combination of the language large model and the embedded model is a key technology for embedding damage assessment criteria into the language large model. This technology can use damage assessment criterion information as a private knowledge base to realize a professional damage language large model, which is a complex and systematic engineering project. The damage inference technology based on damage features is designed as follows... Figure 2 As shown.

[0039] Furthermore, integrating the initial damage assessment criteria into the language big model as prior knowledge involves the following steps: First, the initial damage assessment criteria are preprocessed to ensure direct embedding into the corresponding language big model. Specifically, a suitable embedded model is selected for training or fine-tuning, and the initial damage assessment criteria are input into the trained embedded model for feature representation to generate high-quality text representation vectors. At this stage, the model parameters of the embedded model need to be adjusted according to specific requirements, and techniques such as data augmentation and cross-modal fusion are used to improve the embedded model's understanding of specific domain knowledge. Second, the text representation vectors are input into the language big model as prior knowledge. The language big model is then used to clean, filter, and preprocess the text representation vectors (this process may include word segmentation removal, stop word removal, and long text summarization to ensure the quality and consistency of the input data). Finally, the data is stored in the local database of the language big model (i.e., the language big model is deployed locally to ensure efficient system operation and rapid response). In addition, to improve retrieval efficiency, an index structure needs to be built to quickly match and retrieve relevant knowledge content. Understandably, by combining a large language model with an embedded model, this invention can efficiently build specialized large models locally, supporting complex semantic understanding, intelligent question answering, and decision support applications. This approach not only improves the efficiency of knowledge management but also provides highly customized and private solutions for battlefield environments.

[0040] Step S3: Calculate the damage accuracy and screen damage cases based on the target damage level and the manually calibrated building damage level, and use the screened new damage cases to generate new damage assessment criteria.

[0041] In step S3, the correctness of each damage case (which includes multiple damage cases) is judged by comparing the target damage level obtained through language large model inference with the manually labeled building damage levels in the typical target damage case library. Correctly assessed damage cases are used as positive samples, and the proportion of positive samples in all damage cases is calculated to obtain the damage accuracy (i.e., the inference accuracy of the damage level). The correct damage cases are selected as new damage cases, and logical reasoning and rule synthesis are performed using the new damage cases to generate new damage assessment criteria (i.e., the preliminary optimized damage assessment criteria).

[0042] It should be noted that the generation of new damage assessment criteria mainly involves two tasks: summarizing new damage cases and updating the damage assessment criteria. First, the new damage cases are compiled and their accuracy is manually verified to ensure accuracy and reduce the pollution of the knowledge base by noisy data. Then, a language model is used to extract features and perform semantic analysis on the new damage cases, generating high-dimensional vector representations for association and comparison with the content of the initial damage assessment criteria. This process requires the language model to have strong learning and summarizing capabilities, enabling it to better understand and absorb new knowledge points. After completing the knowledge association and comparison, the language model is used to summarize the new damage cases, forming new damage assessment criteria. Finally, the new damage assessment criteria are integrated into the initial damage assessment criteria to ensure their searchability and usability. This entire process not only improves the adaptability of the language model but also enables dynamic updates to the damage assessment criteria, providing more comprehensive support for damage assessment in subsequent complex scenarios. In this way, large language models can efficiently handle unseen cases, and with the continuous accumulation and optimization of knowledge, they can play a greater role in real-world application scenarios.

[0043] Step S4: Based on the new damage assessment criteria and new damage cases, repeat steps S2 to S3 until the damage accuracy no longer increases, and recognize the latest damage assessment criteria as the optimized damage assessment criteria.

[0044] In step S4, based on the new damage cases and new damage assessment criteria generated in step S3, steps S1 to S3 are repeated to obtain multiple optimized building assessment criteria and corresponding damage accuracy. When accuracy is damaged When the damage assessment criteria cease to rise, the optimization is considered complete, resulting in a well-trained large-scale language model and optimized damage assessment criteria. Understandably, this invention iteratively executes steps S1-S3, continuously introducing new expert samples and assessment feedback, and approximating the optimal solution for assessment accuracy through multiple iterations, ultimately forming an assessment criterion system highly adapted to actual combat scenarios (e.g., ...). Figure 3 (As shown).

[0045] After obtaining the trained language model and the optimized damage assessment criterion, the optimized damage assessment criterion needs to be re-embedded into the language model to verify the accuracy of the language model's damage level determination. In the testing phase, the pre- and post-damage remote sensing images and the optimized damage assessment criterion are used as input, and the predicted damage level is output to calculate the accuracy. The testing phase includes the following steps:

[0046] Step S5: Input the remote sensing images before and after the damage into the semantic description model of building damage to obtain the target damage features.

[0047] In step S5, the remote sensing images before and after the damage are input into the trained building damage semantic description model to obtain qualitative damage features, and then the target damage features are obtained based on the qualitative damage features. It should be noted that, to ensure the accuracy of the building damage semantic description, this invention designs a building damage semantic description model based on the RemoteCLIP (Remote Sensing Image-Text Alignment Large Model), image-text fusion, spatiotemporal feature fusion technology, and a Transformer text decoder. This model utilizes the powerful ability of the RemoteCLIP to analyze and understand complex structural information in remote sensing images, obtains HS characteristic information (i.e., damage semantic description information), and generates corresponding text descriptions from visual features. Figure 4 As shown, the semantic description model for building damage comprises three parts: an image and text information extraction module (encoding part), a feature information fusion module, and a text generation module (decoding part). The image and text information extraction module extracts key feature information from the damaged image, including image and text features. The feature information fusion module fuses image and text features at multiple scales to obtain fused feature information. The text generation module generates more accurate and coherent text based on the fused feature information. The image and text information extraction module mainly consists of a large remote sensing image-text alignment model and a LoRA layer. The feature information fusion module consists of image-text fusion and spatiotemporal fusion. The text generation module consists of a multi-head attention mechanism with masking and a Transformer decoder.

[0048] (1) Image and text information extraction module

[0049] The image and text information extraction module mainly consists of a large remote sensing image-text alignment model and LoRA layers. LoRA layers are embedded in the fully connected layers of the large remote sensing image-text alignment model. The large remote sensing image-text alignment model is one of the important technologies for current remote sensing image processing tasks. It can learn more visual features, thereby improving task performance, and it can also "understand" the concept of objects in vision. The large remote sensing image-text alignment model (RemoteCLIP) is an improved version of CLIP (Contrastive Language-Image Pre-training) applied to the field of remote sensing. It is trained through a large number of image-text pairs, learning to map images and text to the same embedding space. The large remote sensing image-text alignment model has made some improvements and optimizations based on CLIP to improve the model's performance and robustness. The main features of the large remote sensing image-text alignment model include: ① Powerful image processing capabilities: capable of processing high-resolution and large-scale remote sensing images; ② Efficient feature extraction: using advanced algorithms to effectively identify and extract key information from images; ③ Automated and real-time analysis: capable of automatically processing large amounts of data and providing real-time analysis. Meanwhile, to make the remote sensing image-text alignment model more suitable for target damage semantic description tasks, this paper uses a LoRA layer to fine-tune the remote sensing image-text alignment model, so that the features extracted by the remote sensing image-text alignment model are more focused on the target damage semantic description task. In addition, the core working principle of embedding the LoRA layer in the fully connected layer of the remote sensing image-text alignment model is shown in Equation (1).

[0050] (1)

[0051] in, The input features of the LoRA layer are represented by A and B, which are two updatable weight parameters, and W represents the pre-trained weights. This represents the weight increment that needs fine-tuning. The dot product of B and A. It is a scaling factor. Represents dot product. Represents matrix multiplication. This represents the input features of the LoRA layer.

[0052] (2) Feature information fusion module

[0053] The feature information fusion module consists of two parts: visual-text fusion and spatio-temporal fusion. It achieves deep collaboration of multimodal and multi-temporal data through a three-level linkage design of "feature alignment - cross-modal interaction - spatio-temporal aggregation". In the visual-text fusion part, the Transformer module and multimodal feature retrieval are used to complete the spatial alignment and semantic information alignment of cross-modal features, and a unified semantic space is constructed using learnable affine transformations. In the spatio-temporal fusion part, a multi-head cross-attention mechanism is used to establish fine-grained associations between modalities, and dynamic information interaction is achieved by combining gated feature filtering and residual fusion. Finally, the Mamba attention mechanism is used to capture long-range dependencies, and the relationship between multi-temporal features and the correspondence rules between semantic information and text information are established through feature weighting and residual connections. This design particularly emphasizes the joint modeling of spatio-temporal features, introducing a dynamic weight allocation mechanism to adaptively adjust the contribution of multimodal feature information, and adopting a multi-scale fusion strategy to simultaneously extract local details and global semantics, enhancing the model's granular perception of images. Furthermore, the working principles of the Transformer module and the Mamba attention mechanism are shown in equations (2) and (3).

[0054] (2)

[0055] (3)

[0056] in,( )represent Module input characteristics, represent Input features of attention mechanisms . Represents a feedforward neural network. Represents the activation function. Represents a fully connected layer. Represents dot product. Represents matrix multiplication. It is a constant. Represents the state space. This represents a convolutional layer.

[0057] (3) Text generation module

[0058] The text generation module is crucial in determining the plausibility of damaged descriptions. It incorporates a multi-head attention mechanism with masking and a Transformer decoder, enabling it to better capture long-range dependencies without suffering from the vanishing or exploding gradient problems common in traditional recurrent neural networks. This allows the Transformer decoder to maintain contextual consistency when generating long text, considering all positions in the input sequence, thus better capturing global information and generating more accurate and coherent text.

[0059] Step S6: Input the target damage characteristics and the optimized damage assessment criteria into the language big model to obtain the target damage level.

[0060] It should be noted that the optimized damage assessment criteria are integrated into the trained language model as prior knowledge, enabling the language model to determine the damage level of buildings based on the qualitative damage characteristics of the optimized criteria. Understandably, compared to existing threshold-based methods for assessing building damage levels, this invention embodies a collaborative mechanism of "data-damage assessment criteria-model," emphasizing both the objectivity of remote sensing imagery and the guiding role of domain knowledge, while leveraging a large model to improve the analysis efficiency and generalization ability for complex scenarios.

[0061] Finally, to verify the feasibility of this technology, this invention utilized a database of over 1000 building case studies to conduct research on optimizing building damage assessment criteria. The optimization results of the building damage assessment criteria are as follows: Figure 5 As shown, in the optimized building damage assessment criteria, for damage level 1, terms such as "minor" and "partial" appear, which are consistent with the characteristics of minor damage; for damage level 2, terms such as "severe" and "multiple" appear, which are consistent with the characteristics of moderate damage; and for damage level 3, terms such as "complete" and "numerous" appear, which are consistent with the characteristics of severe damage. Furthermore, compared to the initial assessment criteria, some new damage characteristic descriptions that can be used as a basis for judging building damage are included, such as: burn marks, and rubble surrounding the building.

[0062] To verify the practicality of the optimized building assessment criteria, this study assessed the damage levels of 4300 pairs of buildings before and after damage, using both the initial and optimized criteria, and compared the results. Based on the optimized criteria, the accuracy of damage level assessment increased from 64.74% to 76.70% in the 4300 pairs of damaged buildings. Furthermore, considering the uneven distribution of damage categories among the 4300 image pairs, 25 image pairs were collected for each damage category, resulting in 100 pairs of damaged buildings for testing. The damage assessment accuracy increased from 68.00% to 88.00%, as shown in the experimental results. Figure 6As shown, the advantages of the optimized assessment criteria are mainly reflected in the following aspects: Damage Level 1: misjudgment rate reduced from 20% to 8%; Damage Level 3: accuracy rate increased from 16% to 84%.

[0063] Based on the same technical concept as the foregoing embodiments, this invention also provides a remote sensing image building damage assessment criterion optimization system, comprising: a data acquisition module for acquiring a typical target damage case library and initial damage assessment criteria, wherein the typical target damage case library includes qualitative damage characteristics of buildings and manually labeled building damage levels; a damage level determination module for inputting the typical target damage case library and initial damage assessment criteria into a language large model to obtain corresponding target damage levels; wherein the initial damage assessment criteria are incorporated into the language large model in the form of prior knowledge, so that the language large model determines the damage level of buildings based on the qualitative damage characteristics of the initial damage assessment criteria; and accuracy calculation and... The case screening module is used to calculate the damage accuracy and screen damage cases based on the target damage level and the manually calibrated building damage level, and to generate new damage assessment criteria using the screened new damage cases; the damage assessment criterion optimization module is used to re-optimize the damage assessment criteria based on the new damage cases and the new damage assessment criteria until the damage accuracy no longer increases, and the latest damage assessment criteria is recognized as the optimized damage assessment criteria; the target damage feature acquisition module is used to input remote sensing images before and after damage into the building damage semantic description model to obtain the target damage features; the target damage level acquisition module is used to input the target damage features and the optimized damage assessment criteria into the language big model to obtain the target damage level.

[0064] Based on the same technical concept as the foregoing embodiments, the present invention also provides a non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the aforementioned method for optimizing the assessment criteria for building damage from remote sensing images.

[0065] In summary, this invention optimizes damage assessment criteria on two main levels: First, it expands the scope of damage assessment criteria. Leveraging the powerful reasoning and inductive capabilities of large language models, the system can extract damage features beyond typical structural damage (such as craters and collapses) from a wealth of expert knowledge and real-world cases. These features include, but are not limited to, the target's burning state, degree of fragmentation and disintegration, and subtle morphological changes such as structural deformation and tilt. Standardized descriptions of these explicit damage indicators significantly enhance the completeness and expressiveness of the assessment criteria. Second, it enables the self-iteration and optimization of damage assessment criteria. This invention first performs a preliminary damage assessment of remote sensing imagery based on existing damage assessment criteria and compares the assessment results with expert-annotated cases, selecting correctly assessed cases as positive samples. Based on this, through logical reasoning and rule synthesis, a preliminary optimized damage assessment criterion is generated. This process is repeated iteratively, continuously introducing new expert samples and assessment feedback, and through multiple iterations, approximating the optimal solution for assessment accuracy, ultimately forming a set of assessment criteria highly adapted to actual combat scenarios. This method not only improves the accuracy and timeliness of damage assessment, but also enhances the adaptability of the assessment system in complex and ever-changing battlefield environments, providing more reliable intelligence support for command and decision-making.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing building damage assessment criteria using remote sensing imagery, characterized in that, include: Step S1: Obtain a typical target damage case library and initial damage assessment criteria. The typical target damage case library includes the qualitative damage characteristics of buildings and manually calibrated building damage levels. Step S2: Input the typical target damage case library and the initial damage assessment criteria into the language big model to obtain the corresponding target damage level; wherein, the initial damage assessment criteria are incorporated into the language big model in the form of prior knowledge, so that the language big model can determine the damage level of the building based on the qualitative damage characteristics of the building according to the initial damage assessment criteria. Step S3: Based on the target damage level and the manually calibrated building damage level, calculate the damage accuracy and screen damage cases, and use the screened new damage cases to generate new damage assessment criteria; Step S4: Based on the new damage assessment criteria and new damage cases, repeat steps S2 to S3 until the damage accuracy no longer increases, and recognize the latest damage assessment criteria as the optimized damage assessment criteria.

2. The method for optimizing building damage assessment criteria using remote sensing imagery as described in claim 1, characterized in that, The initial damage assessment criteria are incorporated into the language model in the form of prior knowledge, including: The initial damage assessment criteria are input into the trained embedded model to obtain the corresponding text representation vector; The text representation vector is input as prior knowledge into the language big model and stored in the local database of the language big model.

3. The method for optimizing building damage assessment criteria using remote sensing imagery as described in claim 1, characterized in that, New damage assessment criteria are generated using the newly selected damage cases, including: By using a large language model, feature extraction and semantic analysis are performed on new damage cases to generate high-dimensional vector representations, which can then be correlated and compared with the initial damage assessment criteria. After completing the correlation and comparison with the initial damage assessment criteria, the new damage cases are summarized using a large language model to form new damage assessment criteria.

4. The method for optimizing building damage assessment criteria using remote sensing imagery as described in claim 1, characterized in that, After obtaining the optimized damage assessment criteria, the optimization method further includes: The remote sensing images before and after the damage are input into the semantic description model of building damage to obtain the target damage features; The target damage characteristics and the optimized damage assessment criteria are input into the language big model to obtain the target damage level.

5. The method for optimizing building damage assessment criteria using remote sensing imagery as described in claim 4, characterized in that, The semantic description model of building damage includes an image and text information extraction module, a feature information fusion module, and a text generation module; the image and text information extraction module is used to extract key feature information from remote sensing images before and after damage, and the key feature information includes image feature information and text feature information. The feature information fusion module is used to fuse image feature information and text feature information at multiple scales to obtain fused feature information; The text generation module is used to generate corresponding text based on fused feature information.

6. The method for optimizing building damage assessment criteria using remote sensing imagery as described in claim 5, characterized in that, The image and text information extraction module includes a large model for remote sensing image and text alignment and a LoRA layer. The formula corresponding to the LoRA layer is: ， in, The input features of the LoRA layer are represented by A and B, which are two updatable weight parameters, and W represents the pre-trained weights. This represents the weight increment that needs fine-tuning. The dot product of B and A. It is a scaling factor. Represents dot product. Represents matrix multiplication. This represents the input features of the LoRA layer.

7. The method for optimizing building damage assessment criteria using remote sensing imagery as described in claim 5, characterized in that, The feature information fusion module includes an image-text fusion part and a spatiotemporal fusion part. In the image-text fusion part, the Transformer module and multimodal feature retrieval are used to complete the spatial alignment and semantic information alignment of cross-modal features, and a unified semantic space is constructed using learnable affine transformations. In the spatiotemporal fusion part, a multi-head cross-attention mechanism is used to establish fine-grained associations between modalities, and dynamic information interaction is achieved by combining gated feature filtering and residual fusion. The Mamba attention mechanism is used to capture long-range dependencies, and the relationship between multi-temporal features and the correspondence rules between semantic information and text information are established through feature weighting and residual connections.

8. The method for optimizing the assessment criteria for building damage from remote sensing images as described in claim 7, characterized in that, The text generation module includes a masked multi-head attention mechanism and a Transformer decoder. The formulas corresponding to the Transformer module and the Mamba attention mechanism are as follows: ， ， in,( ) represents the input features of the Transformer module. for Input features of the attention mechanism; Represents a feedforward neural network. Represents the activation function. Represents a fully connected layer. Represents dot product. Represents matrix multiplication. It is a constant. Represents the state space. This represents a convolutional layer.

9. A system for optimizing building damage assessment criteria based on remote sensing imagery, characterized in that, include: The data acquisition module is used to acquire a typical target damage case library and initial damage assessment criteria. The typical target damage case library includes the qualitative damage characteristics of buildings and manually calibrated building damage levels. The damage level determination module is used to input the typical target damage case library and the initial damage assessment criteria into the language big model to obtain the corresponding target damage level; wherein, the initial damage assessment criteria are incorporated into the language big model in the form of prior knowledge, so that the language big model can determine the damage level of the building based on the qualitative damage characteristics of the building according to the initial damage assessment criteria. The accuracy calculation and case screening module is used to calculate the damage accuracy and screen damage cases based on the target damage level and the manually calibrated building damage level, and to generate new damage assessment criteria using the screened new damage cases; The damage assessment criteria optimization module is used to re-optimize the damage assessment criteria based on new damage cases and new damage assessment criteria until the damage accuracy no longer increases. The latest damage assessment criteria are then recognized as the optimized damage assessment criteria.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the method for optimizing the assessment criteria for building damage from remote sensing images as described in any one of claims 1 to 8.

Images