Disaster site multi-dimensional image cooperative processing method based on unmanned aerial vehicle-satellite cooperation

Abstract

The invention discloses a disaster site multi-dimensional image cooperative processing method based on unmanned aerial vehicle-satellite cooperation. Acquiring multi-source heterogeneous image data, performing time-space synchronization preprocessing, and realizing timestamp correction by constructing a communication delay prediction model; then establishing a three-dimensional evaluation system based on time evidence, space evidence and consistency evidence by adopting a boundary evidence intensity driven segmentation processing strategy, and carrying out differential registration processing on the image boundary; then realizing balance between boundary information keeping and continuity under a constraint optimization framework through a boundary continuity constraint propagation mechanism and a sliding window reconstruction technology; and executing multi-dimensional image fusion of quality perception, and outputting a high-precision disaster situation map and a loss evaluation result. According to the method, the technical problems of low image registration precision and'ghost boundary 'generated by boundary fusion under heterogeneous timestamps are solved, and the accuracy and reliability of disaster loss assessment are improved.

Description

Technical Field

[0001] The present invention relates to communication technology, in particular to a collaborative processing method of multi-dimensional images of disaster sites based on UAV-satellite collaboration. Background Art

[0002] Against the backdrop of intensifying global climate change and frequent natural disasters, rapid disaster site assessment and quantitative loss analysis are crucial for rescue decision-making and resource allocation. Multidimensional image collaborative processing technology, integrating low-altitude drone swarms with satellite communication networks, provides multi-angle and multi-scale observation capabilities at disaster sites, overcoming the limitations of single-platform observation. Satellite imagery offers the advantages of wide coverage and stable revisit cycles, while drone swarm imagery offers high resolution and flexible acquisition. This fusion of the two not only complements their respective strengths but also provides accurate data support for disaster loss assessment, rescue route planning, and resource scheduling, with significant practical implications for improving the efficiency of disaster emergency response and the success rate of rescue operations.

[0003] Current multi-platform image fusion technology mainly focuses on traditional image registration and fusion algorithm optimization. In the field of image registration, researchers mainly use methods based on feature point matching, such as SIFT, SURF and other algorithms to perform geometric registration of multi-source images, and achieve spatial alignment between images from different platforms by extracting scale-invariant feature descriptors. In terms of image fusion, existing technologies mostly use traditional fusion methods such as wavelet transform, principal component analysis, and Laplace pyramid to achieve the integration of image information through multi-scale decomposition and coefficient fusion. In terms of communication network optimization, research focuses on multi-platform collaborative communication protocol design and network topology optimization, ensuring the reliable transmission of image data by establishing a stable data transmission link. Some research also involves image processing methods based on deep learning, using convolutional neural networks to achieve automatic image registration and fusion.

[0004] However, existing technologies face specific technical challenges when processing heterogeneous, time-stamped images from disaster sites. First, traditional feature point matching algorithms assume that the local texture features of the same feature remain relatively stable across images taken at different times. However, at disaster sites, rapid changes in feature state (such as building collapse and vegetation damage) can cause feature descriptors to become ineffective, significantly reducing registration accuracy. Second, existing image fusion methods generally employ a globally unified processing strategy, failing to distinguish between areas experiencing real changes and those experiencing only differences in imaging conditions. This leads to "ghost boundaries" during boundary fusion, where real feature changes are incorrectly smoothed or false change information is retained, severely impacting the accuracy of subsequent damage assessments. The root cause of these technical issues lies in the lack of specialized processing mechanisms tailored to the characteristics of disaster scenarios. Existing methods fail to fully consider the impact of temporal factors on the image registration and fusion processes. Summary of the Invention

[0005] The purpose of the invention is to provide a method for collaborative processing of multi-dimensional images of disaster sites based on drone-satellite collaboration, in order to solve at least one technical problem existing in the prior art.

[0006] The technical solution is a collaborative processing method for multi-dimensional images of disaster sites based on UAV-satellite collaboration, which includes the following steps:

[0007] Acquire multi-source heterogeneous image data, perform correction and communication delay compensation, establish a unified spatiotemporal reference system, and generate a spatiotemporal corrected image dataset;

[0008] Based on a spatiotemporally corrected image dataset, a boundary evidence strength synthesis model is applied to decompose image boundaries into processing segments and a differentiated registration strategy is implemented based on the evidence strength. This eliminates discontinuous areas in the segmented processing and generates accurately registered image pairs and boundary quality assessment reports.

[0009] Based on this, multi-dimensional image fusion processing is performed to generate a fused situation image, quantitatively analyze the loss area, and obtain the fused situation image and disaster loss assessment results.

[0010] Furthermore, a comprehensive model of boundary evidence strength is applied to implement a differentiated registration strategy, including:

[0011] Select the image pair with the largest time interval from the spatiotemporal correction image dataset as the image pair to be registered;

[0012] A comprehensive boundary evidence strength model is constructed and used to evaluate the image pairs to be registered and generate a boundary evidence strength map.

[0013] The continuous boundary is decomposed into processing segments based on the boundary evidence strength map, and differential fusion is implemented to generate segmented processing results.

[0014] Furthermore, the boundary evidence strength comprehensive model is applied to evaluate the image pairs to be registered and generate a boundary evidence strength map, including:

[0015] Perform edge detection on the image pair to be registered, extract boundary pixels whose gradient amplitude is greater than the threshold, and establish a boundary pixel set and gradient feature record table;

[0016] Calculate the temporal evidence strength of each boundary point in the boundary pixel point set based on the temporal difference between images and generate a temporal evidence strength graph;

[0017] Calculate the gradient difference of each boundary point in the two images according to the gradient feature record table to generate a spatial evidence strength map;

[0018] A window is established around each boundary point to analyze the spatial distribution pattern of boundary changes, calculate the consistency of neighborhood changes, and generate a consistency evidence strength map;

[0019] The time, space and consistency evidence strength maps are weightedly fused to form a boundary evidence strength map.

[0020] Furthermore, the continuous boundary is decomposed into processing segments based on the boundary evidence strength map and differential fusion is performed, including:

[0021] Perform connected domain analysis on the boundary evidence strength map to identify continuous boundary areas, divide them into equal-length segments, and generate a boundary segment identification table;

[0022] Combined with the boundary segment identification table, the average evidence strength value of each boundary segment is calculated and compared with the threshold, and the boundary segments are divided into high evidence strength segments, medium evidence strength segments and low evidence strength segments to form a boundary segment classification table;

[0023] For high evidence intensity segments, the pixel values ​​at the boundary of the new image are replaced. For medium evidence intensity segments, the evidence intensity weighted fusion is used. For low evidence intensity segments, Poisson gradient fusion is performed. The intermediate results of segmentation processing are generated and spatially spliced ​​to obtain the segmentation processing results.

[0024] Furthermore, discontinuous areas in the segmentation process are eliminated to generate a precise registration image pair and a boundary quality assessment report, including:

[0025] Through the constraint propagation network between boundary segments, the processing strategy differences of adjacent segments in the segmentation processing results are detected. When the difference exceeds the threshold, the processing method is adjusted to generate the constraint-adjusted boundary;

[0026] The sliding window reconstruction technique is used to optimize the boundary after constraint adjustment, and boundary continuity is achieved under the constrained optimization framework to obtain accurately registered image pairs and boundary quality assessment reports.

[0027] Furthermore, the calculation of the temporal, spatial and consistency evidence strength maps includes:

[0028] Perform edge detection on the image pair to be registered, extract boundary pixels with gradient magnitude greater than the threshold, calculate the gradient direction angle of each boundary point, and establish a boundary pixel set and gradient feature record table;

[0029] Based on the time difference between images, the time evidence strength of each boundary point in the boundary pixel point set is calculated as the ratio of the time difference to the time threshold, and a time evidence strength map is generated;

[0030] Read the gradient feature record table, calculate the ratio of the gradient difference between the two images and the maximum gradient value of each boundary point as the spatial evidence strength, and generate a spatial evidence strength map;

[0031] An n×n pixel window is established around each boundary point, and the ratio of the number of pixels in the window that are consistent with the neighborhood change direction to the total number of pixels is calculated as the consistency evidence strength to generate a consistency evidence strength map, where n is a preset value.

[0032] Furthermore, the construction process of the comprehensive model of boundary evidence strength includes:

[0033] Constructing a temporal evidence evaluation function based on the temporal differences in the image pair to be registered;

[0034] Calculate image gradient differences and construct a spatial evidence evaluation function based on them;

[0035] Analyze the coherence of local change patterns and construct a consistency evidence evaluation function based on it;

[0036] Weights are configured for the temporal evidence evaluation function, spatial evidence evaluation function, and consistency evidence evaluation function, and weighted fusion is performed to form a comprehensive model of boundary evidence strength.

[0037] Furthermore, the constraint-adjusted boundary is generated, including:

[0038] Based on the segmentation processing results, an adjacency relationship graph between boundary segments is constructed, the difference in processing strategies between spatially adjacent boundary segments is calculated, and an inter-segment constraint relationship graph is generated;

[0039] A continuity penalty function is designed based on the inter-segment constraint relationship graph. When the difference in processing strategies exceeds a threshold, a constraint adjustment mechanism is triggered and a constraint adjustment instruction table is generated.

[0040] The segmentation processing results are locally reprocessed according to the constraint adjustment instruction table, the processing strategy of high-difference adjacent segments is adjusted to the intermediate category processing method, and the constraint-adjusted boundary is output.

[0041] Furthermore, the sliding window reconstruction technology is used for optimization, including:

[0042] A sliding window with a length of a pixels is established for the boundary after constraint adjustment. The window step size is set to b pixels so that each boundary segment is completely covered, and a sliding window position table is generated; a and b are preset values;

[0043] Based on the sliding window position table and the constraint-adjusted boundary, a constrained optimization objective function including evidence preservation term, smoothing term and original information preservation term is established for each sliding window to solve the optimal boundary position and generate a window optimization result set;

[0044] The overlapping boundary areas in the window optimization result set are fused using the distance weighted average method to obtain accurately registered image pairs and boundary quality assessment reports.

[0045] Furthermore, communication delay compensation includes:

[0046] Build a delay prediction model based on the network status parameters of the communication link and calculate the delay compensation value of each transmission link;

[0047] Apply the delay compensation value to retroactively correct the image transmission timestamp information to eliminate the systematic deviation of the timestamp caused by the delay difference of different communication links, and generate a correction timestamp record table;

[0048] Based on the correction timestamp record table, a unified time axis based on the earliest image time is established, the image sequence is reorganized, and the spatiotemporal correction image dataset is output.

[0049] Furthermore, constructing a delay prediction model based on the communication link network state parameters includes:

[0050] Extract the historical delay data series of each communication link from the communication link network status parameters, calculate the delay mean, standard deviation, and change trend by time window group, identify the periodicity and mutation characteristics of the delay pattern, and generate a delay pattern feature table;

[0051] Based on the delay pattern feature table, we analyze the correlation between bandwidth utilization and delay and establish a piecewise linear relationship model. We also establish a model for the impact of network load on delay and output a delay impact factor model.

[0052] A comprehensive delay prediction function is constructed by combining the delay pattern feature table and the delay impact factor model. The prediction accuracy is verified through a sliding window and the parameters are adaptively adjusted to output the delay prediction model.

[0053] Furthermore, a multi-dimensional image fusion process is performed in combination with the precisely registered image pairs and the boundary quality assessment report, including:

[0054] Based on the boundary positioning accuracy information in the boundary quality assessment report, quality-aware wavelet fusion is performed on the precisely registered image pair. A detail-preserving fusion strategy is used in areas with high boundary quality, while a spectrum-preserving fusion strategy is used in areas with low boundary quality to generate a fused situation image.

[0055] Combined with the boundary quality assessment report, the quality-aware area calculation of the ground feature change area in the fused situation image is performed. Loss indicators such as the area of ​​collapsed buildings, the length of blocked roads, and the extent of vegetation damage are calculated by classification, and a classified loss statistics table is established.

[0056] A loss degree grading evaluation system is established based on the classified loss statistics table and boundary quality information, and the disaster loss assessment results and treatment quality verification report are output.

[0057] Furthermore, performing quality-aware wavelet fusion based on the boundary quality assessment report includes:

[0058] Detect the sharply changing boundary regions in the precisely registered image pairs, use Daubechies wavelet to maintain spectral fidelity in the smooth regions, use Biorthogonal wavelet to reduce the ringing effect in the sharply changing regions, and establish an adaptive wavelet basis selection map;

[0059] Perform multi-layer wavelet decomposition on the precisely registered image pair based on the adaptive wavelet basis selection map, decompose the image into low-frequency approximation coefficients and high-frequency detail coefficients, and generate a multi-scale wavelet coefficient set;

[0060] Based on the boundary positioning accuracy information in the boundary quality assessment report, the local fusion weight is calculated, and the dynamic weight is applied to the multi-scale wavelet coefficient set to perform coefficient fusion to generate the quality-aware fusion coefficient.

[0061] The quality-aware fusion coefficients are reconstructed by inverse wavelet transform and boundary-preserving constraints are introduced to output the fused situation image.

[0062] Furthermore, the quality-aware area calculation based on the boundary quality assessment report includes:

[0063] Identify the areas of ground feature changes in the fused situation image, use precise pixel counting for areas with high boundary quality, introduce uncertainty coefficients for areas with medium boundary quality, and adjust the area to generate a quality-aware change area map;

[0064] Based on the quality perception change area map, the collapsed building area, road blockage length, and vegetation damage range are classified and calculated. Each loss indicator is associated with a corresponding boundary quality score for reliability assessment, and a classified loss statistics table is established.

[0065] According to the classification loss statistics table and boundary quality information, the degree of loss is divided into four levels: slight loss, moderate loss, severe loss and uncertain loss. The area, proportion and confidence level of each level of loss are calculated, and the disaster loss assessment results are output.

[0066] The beneficial effect is solved by constructing a segmentation processing mechanism driven by boundary evidence strength.

[0067] Specifically, we no longer rely on fixed SIFT feature descriptors for global unified matching. Instead, we establish a three-dimensional evaluation system based on temporal evidence, spatial evidence, and consistency evidence. We dynamically calculate the change evidence strength of each boundary point and decompose the boundary into 3-5 pixel processing segments based on the evidence strength value. For high evidence strength segments, the new image boundary information is directly retained to maintain the true change. For medium evidence strength segments, evidence strength weighted fusion is used, and for low evidence strength segments, traditional gradient fusion is performed. This differentiated processing strategy effectively avoids the problem of feature matching failure caused by ground object changes and significantly improves the registration accuracy of heterogeneous timestamp images.

[0068] The solution is achieved through boundary continuity constraint propagation and sliding window boundary reconstruction technology.

[0069] On the one hand, by constructing an adjacency relationship graph between boundary segments and designing a continuity penalty function to detect the strategy differences between adjacent processing segments, when the difference exceeds a threshold, a constraint adjustment mechanism is triggered, and the processing strategy of high-discrepancy adjacent segments is adjusted toward the intermediate category, ensuring the consistency of local processing decisions. On the other hand, a 9-pixel sliding window is used to cover three boundary segments, and a constrained optimization objective function is established, which includes evidence preservation terms, smoothing terms, and original information preservation terms. The optimal boundary position is solved by the gradient descent algorithm, and boundary continuity is achieved while maintaining high-evidence change information. This eliminates the boundary discontinuity and "ghost boundary" problems caused by segmented processing, ensuring the accurate identification of real feature change boundaries and the precision of loss assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a flow chart of the present invention.

[0071] Figure 2 It is a flow chart of the present invention using the boundary evidence strength comprehensive model to implement the differentiated registration strategy.

[0072] Figure 3 It is a flow chart of the present invention for generating a boundary evidence strength map.

[0073] Figure 4 The present invention is a flow chart of decomposing continuous boundaries into processing segments and implementing differential fusion based on the boundary evidence strength graph.

[0074] Figure 5 It is a flow chart of the present invention for generating accurate registration image pairs and boundary quality assessment reports. DETAILED DESCRIPTION

[0075] like Figure 1 As shown, a method for collaborative processing of multi-dimensional images of disaster sites based on UAV-satellite collaboration is provided, comprising:

[0076] S1. Acquire satellite multispectral image data and drone swarm multisensor image data, synchronously collect network status parameters and transmission timestamp information of each communication link, apply geometric correction and radiation correction technology to perform image standardization processing, establish a unified geographic coordinate system, and generate standardized multi-source image data sets and communication link status parameter tables as the data basis for subsequent spatiotemporal synchronization processing.

[0077] S2. Read the communication link status parameter table, build a delay prediction model based on the network status history, calculate the delay compensation value of each transmission link, perform retroactive correction on the timestamp information in the standardized multi-source image dataset, establish a unified space-time reference system that takes communication delay into account, and output the space-time corrected image dataset to support accurate image registration processing.

[0078] S3. Process spatiotemporally corrected image datasets, apply boundary evidence strength assessment mechanisms to identify areas of ground feature change, construct a comprehensive strength model based on temporal evidence, spatial evidence, and consistency evidence, decompose image boundaries into multiple processing segments, and implement differentiated registration strategies based on evidence strength. Eliminate discontinuities caused by segmentation through boundary continuity constraint propagation and sliding window reconstruction techniques, and generate precisely registered image pairs and boundary quality assessment reports to address registration accuracy issues under heterogeneous timestamps.

[0079] S4. Based on the precise registration of image pairs and boundary quality assessment reports, perform multi-dimensional image fusion processing to generate a high-precision disaster scene situation map. Combine boundary quality information to conduct quantitative analysis of the loss area. Synchronously output the fused situation image, disaster loss assessment results, and processing quality verification report. Establish a credibility evaluation system for the fusion results to provide reliable data support for disaster relief decision-making.

[0080] Specifically, a method for collaborative processing of multi-dimensional images of disaster sites based on drone-satellite collaboration by integrating a low-altitude drone swarm with a satellite communication network is provided.

[0081] S1: Multi-source heterogeneous data acquisition and synchronous preprocessing

[0082] S11: Acquire satellite multispectral image data, including visible light, near infrared, shortwave infrared band information and corresponding imaging timestamps, synchronously obtain satellite orbit parameters and imaging geometry information, and generate raw satellite image data and satellite imaging parameter records.

[0083] S12: Collect image data from the RGB cameras, thermal infrared sensors, and multispectral cameras carried by the drone swarm, record the GPS position, flight attitude, and image acquisition timestamp of each drone, establish a correspondence between drones and image data, and output the original drone image dataset and drone status record table.

[0084] S13: Monitor the network status of satellite communication links, direct links between drones, and ground base station relay links in real time, collect bandwidth occupancy, signal strength, data packet transmission delay, and packet loss rate parameters of each link, establish network status monitoring records indexed by timestamps, and generate a communication link status parameter table.

[0085] S14: Perform geometric correction on the original satellite image data and the original UAV image dataset to eliminate sensor distortion and imaging geometric errors, uniformly project them into the WGS84 coordinate system, perform radiometric correction to eliminate atmospheric effects and sensor response differences, and output a standardized multi-source image dataset as the basic data for subsequent processing.

[0086] S2: Heterogeneous Timestamp Correction and Communication Delay Compensation

[0087] S21: Read the historical communication link status parameter table, analyze the historical delay pattern of each communication link, and construct the delay prediction function τ pred = α×τ hist + β×BW current + γ×Load current , where τ hist is the historical delay average, BW current is the current bandwidth usage, Load current Build a delay prediction model for the current network load.

[0088] S22: Apply the delay prediction model to calculate the transmission delay compensation value of each image in the standardized multi-source image dataset, and perform back-correction on the image receiving timestamp T corrected = T received - τ pred , it eliminates the systematic deviation of timestamps caused by the delay differences of different communication links and generates corrected image timestamp records.

[0089] S23: Based on the corrected timestamp records, a unified time axis based on the earliest image time is established, the image sequences in the standardized multi-source image dataset are reorganized to ensure that the timestamps reflect the actual image acquisition time relationship, and the spatiotemporally corrected image dataset is output.

[0090] S3: Multidimensional Image Registration Driven by Boundary Evidence Strength

[0091] S31: Read the spatiotemporal correction image dataset, select the satellite image and drone image with the largest time interval as the registration object, calculate the time difference Δt between the two images, identify the overlapping area and extract the initial feature points, and establish the image pair to be registered and the time difference parameters.

[0092] S32: For each boundary pixel in the image pair to be registered, a boundary evidence strength evaluation mechanism is constructed: the time evidence S is calculated. time = |Δt| / T threshold , spatial evidence S spatial = grad I1 - grad I2 (gradient difference), consistency evidence S consist= The coherence score of the local change pattern, obtained by weighted fusion S total = w1×S time + w2×S spatial + w3×S consist The comprehensive evidence strength is obtained and a boundary evidence strength map is generated. grad is the gradient operator.

[0093] S33: Based on the boundary evidence strength map, the continuous boundary is decomposed into processing segments of length 3-5 pixels, and the evidence strength value S of each segment is used to calculate the boundary. total Implement differentiated treatment strategies: High intensity segment (S total >0.7) completely retains the boundary information of the new image, and the medium intensity segment (0.3≤S total ≤0.7) perform weighted fusion of evidence strength, low-intensity segment (S total <0.3) Use traditional gradient fusion to output segmented processing results.

[0094] S34: In view of the strategy differences between adjacent processing segments in the segmented processing results, a constraint propagation network is established between boundary segments. A continuity penalty function is designed to detect the degree of difference in the processing strategies of adjacent segments. When the difference exceeds a threshold, the processing method of the boundary segment is adjusted to ensure the consistency of local boundary processing decisions and generate a constraint-adjusted boundary.

[0095] S35: Apply sliding window boundary reconstruction technology to the constraint-adjusted boundary, using a 9-pixel-width sliding window to cover the three boundary segments. Under the constrained optimization framework, find the optimal boundary position that maintains high-evidence segment information while meeting the boundary continuity requirement. Eliminate boundary jumps and breaks through iterative optimization, and output a precisely registered image pair and a boundary quality assessment report.

[0096] S4: Collaborative integration and result output

[0097] S41: Reads precisely registered image pairs and performs multi-scale image fusion based on the registration accuracy. This optimally combines the spectral information of satellite images with the spatial detail information of drone images. Wavelet transform domain fusion technology is used to maintain spectral fidelity and spatial detail to generate a fused situation image.

[0098] S42: Combined with the boundary positioning accuracy information in the boundary quality assessment report, quantitative analysis is performed on the area of ​​ground feature changes in the fused situation image. Disaster loss indicators such as the area of ​​collapsed buildings, the length of blocked roads, and the range of vegetation damage are calculated. A loss degree grading evaluation system is established to output the disaster loss assessment results.

[0099] S43: Based on the boundary quality assessment report and the error propagation analysis during the fusion processing, the confidence and uncertainty range of the fused situation image and disaster loss assessment results are calculated. A quality evaluation index system including registration accuracy, fusion fidelity, and loss assessment reliability is established. A processing quality verification report is generated to provide a credible reference for decision-making.

[0100] According to one aspect of the present application, S32: Boundary evidence strength assessment mechanism, specifically:

[0101] S321: Read the image pair to be registered and the time difference parameters, perform Sobel operator edge detection on the two images, extract boundary pixels whose gradient amplitude is greater than the threshold Th_grad, calculate the gradient direction angle of each boundary point, establish the correspondence between the spatial coordinates of the boundary point and the gradient feature, and generate a boundary pixel set and a gradient feature record table.

[0102] S322: Calculate the temporal evidence strength S for each boundary point in the boundary pixel set based on the Δt value in the time difference parameter. time = |Δt| / T threshold , where T threshold is the time threshold parameter set according to the disaster type (fire scenario T threshold = 30min, flood scenario T threshold =60min), when Δt exceeds the threshold S time A value approaching 1 indicates a high probability of change, generating a time-based evidence strength plot.

[0103] S323: Read the gradient feature record table and calculate the gradient difference S of each boundary point in the two images. spatial = |grad I1(x,y) – grad I2(x,y)| / max(grad I1(x,y), grad I2(x,y)), where gradI1 and grad I2 are the gradient amplitudes of the two images at the (x,y) position, respectively. When the gradient difference exceeds 0.5, it is considered that a significant spatial change has occurred. A spatial evidence strength map reflecting the degree of spatial change of the boundary points is constructed.

[0104] S324: Combine the boundary pixel points and build a 7×7 pixel window around each boundary point. Analyze the spatial distribution pattern of boundary changes within the window and calculate the consistency evidence S consist = (N consistent -N isolated ) / N total , where N consistent is the number of pixels that are consistent with the neighborhood change direction, N isolated is the number of isolated change pixels, N totalis the total number of pixels in the window, forming a consistency evidence strength map.

[0105] S325: Read the time evidence strength map, space evidence strength map and consistency evidence strength map, and use weighted fusion S total = w1×S time + w2×S spatial + w3×S consist Calculate the comprehensive evidence strength, where the weight parameters w1=0.4, w2=0.4, w3=0.2 are preset according to the characteristics of the disaster scene, and S total The values ​​are normalized to the interval [0,1] and a boundary evidence strength map is output.

[0106] By building a three-dimensional comprehensive evaluation system based on temporal evidence, spatial evidence, and consistency evidence, we have broken through the fundamental limitation of the traditional SIFT feature matching algorithm, which is the failure of feature descriptors when the ground objects at the disaster site change rapidly. This technological innovation transforms the abstract boundary change judgment into a specific numerical evidence strength calculation, in which the temporal evidence S time = |Δt| / T threshold Directly related to the difference between the time of disaster occurrence and the time of image acquisition, spatial evidence S spatial = |gradI1 -gradI2| / max(gradI1, gradI2) quantifies the degree of abrupt change in gradients on both sides of the boundary, and the consistency evidence S consist =(N consist ent - N isolated ) / N total The spatial coherence of the local variation pattern is evaluated. The three are fused by weight S total = 0.4×S time + 0.4×S spatial + 0.2×S consist A comprehensive judgment is formed, enabling the system to accurately identify areas of real ground changes and areas of difference in imaging conditions. In earthquake disaster scenarios, the boundary positioning accuracy is improved from the 3.2-pixel error of traditional methods to 0.8-pixel error, an accuracy improvement of 75%, effectively solving the key problem of registration failure of existing technologies in the complex environment of disaster sites.

[0107] According to one aspect of the present application, S33: Segment boundary processing strategy, specifically:

[0108] S331: Read the boundary evidence strength map, apply the connected domain analysis algorithm to identify continuous boundary regions, divide each continuous boundary region into equal length segments of 3-5 pixels, ensure that each segment contains enough spatial information for independent processing and decision making, assign a unique identifier to each boundary segment, and generate a boundary segment identification table.

[0109] S332: Calculate the average evidence strength value S for each boundary segment based on the boundary segment identification table and the boundary evidence strength map. seg = (1 / N)∑S total (i), where N is the number of pixels in the segment, according to S seg The boundary segments are divided into three categories: high evidence strength segment (S seg >0.7), moderate evidence strength (0.3≤S seg ≤0.7), low evidence strength segment (S seg <0.3), and establish a boundary segment classification table.

[0110] S333: Read the boundary segment classification table and the image pair to be registered, and implement differentiated processing for boundary segments of different categories: the high evidence strength segment directly adopts the boundary pixel value I of the newer image. new (x,y) replaces the registration result, and the medium evidence strength segment adopts the evidence strength weighted fusion I fused (x,y) = S seg ×I new (x,y) + (1-S seg )×I old (x,y), low evidence strength segments perform traditional Poisson gradient fusion gradI fused = αgradI new + (1-α)gradI old , generate segmented processing intermediate results.

[0111] S334: spatially stitch the intermediate results of the segmentation processing, combine the processing results of each boundary segment into a complete image according to the spatial position information in the boundary segment identification table, and detect whether there is a pixel value jump exceeding the threshold Th at the stitching boundary jump =20, record the jump position and amplitude information, and output the segment processing results and boundary jump record table.

[0112] By decomposing continuous boundaries into independent processing segments of 3-5 pixels and implementing differentiated processing strategies based on the strength of evidence, the technical defect of traditional image fusion methods that use global unified processing and result in the loss of real change information is fundamentally changed. This innovation divides boundary segments into three categories according to the strength of evidence: high evidence strength segment (S seg >0.7) Directly use the new image boundary information I new Replace with complete preservation of land feature changes, medium evidence strength segment (0.3≤S seg ≤0.7) using evidence strength weighted fusion I fused = S seg ×I new + (1-S seg )×I oldAchieve progressive information fusion, low evidence strength segment (S seg <0.3) using traditional Poisson gradient fusion to eliminate the effects of imaging differences. This differentiation strategy enables the system to maintain true ground feature changes while effectively eliminating false boundaries. This improved the accuracy of damage assessment in identifying collapsed building areas from 67% with traditional methods to 89.3%, a 33.2% improvement. This completely addresses the core issue of "ghost boundaries" generated by traditional fusion methods, which can lead to inaccurate disaster loss assessments. This provides a reliable quantitative analysis foundation for disaster relief decision-making.

[0113] According to one aspect of the present application, S34: inter-boundary segment constraint propagation, specifically:

[0114] S341: Read the boundary jump record table and the boundary segment classification table, build an adjacency relationship graph between boundary segments, establish connecting edges for spatially adjacent boundary segments, and calculate the difference in processing strategies between adjacent segments D_diff = |category i - Category j |, where categories are coded with numerical values ​​(high = 3, medium = 2, low = 1). When D_diff ≥ 2, it is considered that there is a significant treatment difference, and an inter-segment constraint relationship diagram is established.

[0115] S342: Based on the inter-segment constraint relationship graph, design the continuity penalty function P penalty = exp(D_diff×λ)×|I i -I j |, where λ is the penalty coefficient, I i and I j is the pixel value at the boundary of adjacent segments, when P penalty Exceeding the threshold Th penalty The constraint adjustment mechanism is triggered when the high-difference adjacent segments are processed towards the middle category, and a constraint adjustment instruction table is generated.

[0116] S343: Locally reprocess the segmentation processing results according to the constraint adjustment instruction table, and adopt the revised processing strategy for the boundary segments that need adjustment: the original high-evidence segment is adjusted to the medium-evidence segment processing method, and the original low-evidence segment is adjusted to the medium-evidence segment processing method. The medium-evidence segment processing method remains unchanged to ensure the continuity of the processing strategy between adjacent segments, and the boundary after constraint adjustment is output.

[0117] By constructing a boundary segment constraint propagation network and designing a continuity penalty function P penalty = exp(D_diff×λ)×|I i -I j|, innovatively solves the problem of inconsistent processing strategies for adjacent segments that inevitably result from segmented processing. This technology calculates the difference in processing strategies for adjacent segments D_diff = |category i - Category j When the difference exceeds a threshold, a constraint adjustment mechanism is triggered, adjusting the processing strategy of high-difference adjacent segments toward the intermediate category, ensuring spatial coherence of local processing decisions. This constraint propagation mechanism enables the system to maintain true change information in high-evidence areas while dynamically adjusting inter-segment constraint relationships to avoid sudden changes in processing results in adjacent areas. This reduces the rate of change in boundary curvature by 60% in reconstructing building boundaries at disaster sites, effectively eliminating boundary jumps and discontinuities caused by segmented processing. This innovation ensures that the fusion results maintain the authenticity of ground feature changes while meeting visual continuity requirements, providing a high-quality image foundation for subsequent quantitative loss analysis and significantly improving the reliability and usability of disaster assessment results.

[0118] According to one aspect of the present application, S35: Sliding window boundary reconstruction, specifically:

[0119] S351: Read the constraint-adjusted boundary, establish a sliding window with a length of 9 pixels, set the window step size to 3 pixels to ensure that each boundary segment is completely covered, each window contains 3 consecutive boundary segments, establish a local coordinate system for each window, and generate a sliding window position table and a window content mapping table.

[0120] S352: For each boundary point in the sliding window, establish the constraint optimization objective function F opt = α × E evidence +β×E smooth + γ×E preserve , where E evidence To ensure that high-evidence-strength information is not lost, E smooth To ensure the continuity of the boundary for the smooth term, E preserve To prevent excessive modification of the original information preservation items, the weight parameters α=0.5, β=0.3, and γ=0.2 are used to build a window optimization model.

[0121] S353: Apply the gradient descent algorithm to solve the window optimization model, iteratively update the pixel positions and intensity values ​​of the boundary points within the window, calculate the change ΔF of the objective function value in each iteration, and consider convergence when ΔF < ε_converge = 0.001. Record the optimized boundary point coordinates and pixel values, complete the optimization process for all sliding windows, and generate a set of window optimization results.

[0122] S354: Fusing the overlapping boundary regions in the window optimization result set using the distance weighted average method I final(x,y) = Σ(w i ×I i (x,y)) / Σw i , where w i The inverse weight of the distance from the point (x, y) is used to eliminate the inconsistency of processing between windows, calculate the positioning accuracy and continuity index of the final boundary, and output the precise registration image pair and boundary quality assessment report.

[0123] By establishing a constrained optimization objective function F that includes evidence preservation terms, smoothing terms, and original information preservation terms opt =α×E evidence + β×E smooth + γ×E preserve , and a reconstruction method using a 9-pixel sliding window covering 3 boundary segments was used to achieve boundary continuity reconstruction under the constrained optimization framework. This technical innovation transforms the boundary reconstruction problem into a multi-objective constrained optimization problem, and iteratively solves the optimal boundary position through the gradient descent algorithm, so that the reconstructed boundary maintains the real change information of the high-evidence intensity segment and satisfies the overall continuity constraint. The sliding window mechanism ensures that each boundary segment is completely covered, and the distance weighted average fusion I final (x,y) = Σ(w i ×I i (x,y)) / Σw i This method eliminates inconsistencies in inter-window processing. In reconstructing the boundaries of earthquake-damaged buildings, it reduced the standard deviation of boundary positioning from 2.1 pixels to 0.6 pixels, improving positioning accuracy by 71% while retaining 94% of true change information. This method completely resolves the technical contradiction between maintaining boundary information and achieving continuity in traditional methods, providing a high-precision image registration foundation for accurate loss assessment at disaster sites.

[0124] According to one aspect of the present application, S21: delay prediction model construction, specifically:

[0125] S211: Read the communication link status parameter table, extract the historical delay data series of each communication link, group the data according to the time window of 10 minutes, and calculate the delay mean τ in each time window mean , standard deviation τ std and change trend slope τ ,Identify the periodicity and mutation characteristics of the delay pattern and generate a delay pattern feature table.

[0126] S212: Analyze bandwidth usage BW based on the delay pattern feature table current Correlation with delay τ, establish piecewise linear relationship model: when BW current <0.6 BW = k1×BWcurrent + b1, when BW current ≥0.6 when τ BW = k² × BW current + b2, fitting parameters k1, b1, k2, b2 by least squares method, and establishing network load Load current Impact model on delay τ Load = k3×Load 2 current + k4×Load current + b3, output delay impact factor model.

[0127] In another embodiment of the present application, S212 may also be:

[0128] Based on the delay pattern feature table, a multi-layer delay prediction model based on queuing theory is constructed, and an M / G / 1 queuing system is established to describe the delay characteristics of the communication link: the service delay τ is calculated. service = L packet / BW available , where L packet is the packet length, BW available is the available bandwidth; establish the queuing delay model τ_queue = λ×E[S 2 ] / (2×(1-ρ)), where λ is the packet arrival rate, E[S 2 ] is the second-order moment of service time, ρ = λ×E[S] is the system utilization; the model parameters are dynamically adjusted through the Kalman filter: the state transfer equation X(k+1) = A×X(k) + B×U(k) + W(k), the observation equation Y(k) = C×X(k) + V(k), where X(k) contains the historical delay mean, variance and trend components, W(k) and V(k) are process noise and observation noise, respectively; adaptive weights are set according to different communication link types: the satellite link weight w_sat is adjusted based on the orbital period, the UAV link weight w_uav is adjusted based on the moving speed, and the ground link weight w_gnd is adjusted based on the load fluctuation, to generate an adaptive delay prediction model.

[0129] According to one aspect of the present application, S213 may also be:

[0130] Combine the delay pattern feature table and the delay impact factor model to construct a comprehensive delay prediction function τ pred = α×τ hist + β×τ BW + γ×τ Load + δ×trend factor , where α=0.4, β=0.3, γ=0.2, δ=0.1 are weight coefficients, trendfactor It is a trend correction item based on historical data. The prediction accuracy is verified through a sliding window. When the prediction error exceeds 15%, the parameter adaptive adjustment is triggered to generate a delay prediction model.

[0131] By constructing a delay prediction model τ based on historical analysis of network status pred = α×τ hist + β×τ BW + γ×τ Load + δ×trend factor , and establishes a piecewise linear relationship to describe the correlation between bandwidth occupancy and delay, innovatively solving the problem of image registration timing disorder caused by timestamp asynchrony in multi-platform heterogeneous communication networks. This technology transforms abstract network delay prediction into a specific mathematical model, and through the multi-dimensional fusion of historical delay pattern analysis, bandwidth impact modeling and network load evaluation, it achieves accurate prediction and compensation of the delay of each communication link. Timestamp Correction T corrected =T received - τ pred This method eliminates systematic biases caused by latency differences across different communication links and establishes a unified spatiotemporal reference system. In satellite-UAV collaborative observation scenarios, this method improves timestamp correction accuracy from the traditional ±150ms to ±28ms, an 81% improvement. This ensures consistent time bases during subsequent image registration, provides a reliable timing foundation for collaborative multidimensional image processing with heterogeneous timestamps, and significantly enhances the timeliness and accuracy of situational awareness at disaster sites.

[0132] According to one aspect of the present application, S22: timestamp retroactive correction, specifically:

[0133] S221: Read the delay prediction model and the standardized multi-source image dataset, and identify the transmission link type of each image in the dataset: satellite links are marked as SAT LINK , UAV direct connection marked as UAV DIRECT , the ground relay is marked as GND RELAY , according to the link type, call the corresponding delay prediction sub-model to calculate the delay compensation value τ comp , establish an image delay compensation table.

[0134] In another embodiment of the present application, S221 may also be:

[0135] Read the adaptive delay prediction model and standardized multi-source image dataset to establish a multi-level communication link classification system: the first level classification is the physical link type (satellite Ka band is marked as SAT_KA, satellite Ku band is marked as SAT_KU, and UAV direct connection is marked as UAV). DIRECT , UAV relay marked as UAVRELAY , ground cellular is marked as GND_CELL, ground Wi-Fi is marked as GND_WIFI), the second-level classification is the number of link hops (single hop is marked as SINGLE_HOP, dual hop is marked as DUAL_HOP, and multi-hop is marked as MULTI_HOP), and the third-level classification is the transmission mode (real-time mode RT_MODE, cache forwarding mode SF_MODE, hybrid mode HYBRID_MODE); a dedicated delay prediction sub-model is established for each type of link: the satellite link adopts the orbit period compensation model τ_sat = τ_base + A×sin(2π×t / T_orbit + φ_orbit), the UAV link adopts the Doppler frequency shift compensation model τ_uav = τ_base × (1 + v_relative / c), and the ground link adopts the congestion avoidance model τ_gnd = τ_min × (1 + α×e^(β×Load current )); Establish a link switching detection mechanism, trigger the backup link when the signal quality RSSI < -85dBm or the packet loss rate > 5%, record the switching time and switching delay Δτ_handoff, and update the multi-level link delay compensation table.

[0136] S222: Based on the image delay compensation table, the receiving timestamp T of each image is received Perform back-correction T corrected = T received - τ comp , considering the uncertainty of delay prediction, calculate the confidence interval of the correction timestamp [T corrected - σ τ , T corrected + σ τ ], where σ τ It is the standard error of the delay prediction, provides timestamp reliability evaluation for subsequent processing, and generates a correction timestamp record table.

[0137] S223: Read the correction timestamp record table and evaluate the accuracy of timestamp correction by cross-validation method: select reference images with known real acquisition time, compare the difference between the corrected timestamp and the real timestamp, and calculate the correction accuracy index MAE = (1 / N)Σ|T corrected - T_true|, when MAE>50ms, triggers the correction parameter optimization to ensure that the correction accuracy meets the subsequent registration requirements and outputs the spatiotemporal corrected image dataset.

[0138] In another embodiment of the present application, S22 and S23 may also be:

[0139] S22: Layered Time Synchronization and Frequency Correction

[0140] S221: Establish a hierarchical time synchronization architecture. Set the ground control station as the first-level time server (Stratum 1), which obtains the UTC time reference through a GPS receiver with an accuracy of ±10ns. Set the key UAV nodes as the second-level time server (Stratum 2), which synchronizes with the ground station through the NTP protocol with a synchronization accuracy of ±1ms. The remaining UAVs and satellite nodes are the third-level time clients (Stratum 3), which synchronize with the second-level server to generate a hierarchical time synchronization topology table.

[0141] S222: For each image in the standardized multi-source image dataset, perform multi-step timestamp correction: first perform frequency offset correction T_freq corrected = T_local × (1 + freq_offset / f_nominal), where freq_offset is obtained by comparing the local crystal oscillator with the GPS 1PPS signal; then clock drift correction T_drift is performed corrected =T_freq corrected + drift_rate × (T current - T sync _last), drift_rate is calculated by fitting the latest 100 synchronization records using the least squares method; finally, network delay correction T is performed final = T_drift corrected - (τ_uplink +τ_downlink) / 2, where the uplink and downlink delays are measured through timestamp exchange and an accurate timestamp correction table is established.

[0142] S223: Based on the precise timestamp correction table, establish a timestamp confidence assessment mechanism: calculate the clock synchronization interval Δt sync , when Δt sync < 60s, the confidence level is 0.95. When 60s ≤ Δt sync < 300s Confidence = 0.9 - 0.1×(Δt sync -60) / 240, when Δt sync Confidence = 0.8 when ≥ 300s; combined with the network delay measurement error σ_delay, the confidence level is corrected final = Confidence × exp(-σ_delay 2 / 2σ threshold 2 ), where σ threshold = 10ms; assign confidence intervals [T final - 3σconfidence , T final + 3σ confidence ], generate a confidence-aware timestamp record table.

[0143] S23: Time and Space Reference Unification and Jitter Compensation

[0144] S231: Read the confidence-aware timestamp record table and identify timestamp outliers: Use the 3σ criterion to detect timestamps that deviate from the mean by more than 3 times the standard deviation, and start the secondary correction process for abnormal timestamps: Interpolate and estimate T through the timestamps of adjacent images. i Interpolated = T_prev + (T_next - T_prev) × (sequence position - prev position) / (next position - prev position), weighted correction T is performed based on the image content similarity corrected = w_similarity × T i nterpolated + (1 - w_similarity) × T_original, generates an anomaly-corrected timestamp table.

[0145] S232: Based on the abnormal correction timestamp table, establish a unified time axis for anti-jitter: use sliding median filtering to eliminate short-term jitter T_filtered(i) = median{T(ik), ..., T(i), ..., T(i+k)}, where the window size 2k+1 is adaptively adjusted according to the image acquisition frequency; establish a time base selection algorithm based on minimum confidence loss: T reference = argmax(∑Confidence i × w_proximity i ), select the time point that maximizes the weighted sum of the confidence scores of all images as the unified benchmark; recalculate the time offset of all images relative to the unified benchmark Offset i = T_filtered(i) - T reference ,construct a spatiotemporal correction image dataset considering confidence and jitter compensation.

[0146] According to one aspect of the present application, S14: image normalization processing, specifically:

[0147] S141: Read the original satellite image data and the original UAV image dataset, extract the image EXIF ​​information and sensor parameters, including geometric parameters such as focal length, pixel size, and distortion coefficient, as well as radiation parameters such as radiation calibration coefficient and gain setting, establish the correspondence between the sensor and the image data, and generate a sensor parameter configuration table.

[0148] S142: Based on the sensor parameter configuration table, perform geometric correction processing on each image: use the Brown distortion model r corrected = r×(1 + k1×r 2 + k2×r 4 + k3×r 6 ) to eliminate radial distortion of the lens and apply tangential distortion correction x corrected = x + [2p1×x×y + p2×(r 2 +2x 2 )],y corrected = y + [p1×(r 2 + 2y 2 ) + 2p2×x×y], all images are reprojected into the unified WGS84 coordinate system to generate geometrically corrected image data.

[0149] S143: Perform radiation correction on the geometrically corrected image data, first performing dark current correction DN corrected = DN_raw - DN_dark, then apply radiometric scaling L_radiance = (DN corrected - DN_offset) × Gain factor Finally, the atmospheric scattering effect is eliminated through the atmospheric correction model L_surface = (L_radiance - L_atmosphere) / T_atmosphere, where T_atmosphere is the atmospheric transmittance, to ensure the radiation consistency of images from different sensors and output a standardized multi-source image dataset.

[0150] According to one aspect of the present application, S31: registration object selection, specifically:

[0151] S311: Read the spatiotemporally corrected image dataset, analyze the time interval distribution between images, identify the satellite image and drone image combination with the largest time span, and calculate the time difference Δt for all possible image pairs i j = |T i - T j |, select the image pairs with the largest Δt value and exceeding the preset threshold Δt_min as registration candidates, ensuring that the registration process targets the most challenging temporally heterogeneous scenes and generating a list of candidate registration pairs.

[0152] S312: Perform overlapping area analysis on each image pair in the candidate registration pair list, and calculate the spatial overlap ratio of the two images by geographic coordinate transformation: Overlap_ratio = Area_overlap / min(Area img1,Area i mg2), the overlap rate is required to be greater than 60% to ensure effective registration. At the same time, the feature richness of the overlapping area is analyzed (Feature_density = N_features / Area_overlap), and the image pair with the highest feature density is selected as the final registration object. The image pair to be registered and the time difference parameters are output.

[0153] According to one aspect of the present application, S41: multi-scale image fusion, specifically:

[0154] S411: Read the precisely registered image pairs and detect the sharply changing boundary regions in the images. Traditional wavelet transforms are prone to ringing effects and blurred boundaries in these regions. Therefore, a boundary-sensitive wavelet basis selection mechanism is designed: Daubechies wavelet is used in flat regions to maintain spectral fidelity, and Biorthogonal wavelet is used in sharply changing regions to reduce ringing effects. An adaptive wavelet basis selection map is established.

[0155] S412: Based on the adaptive wavelet basis selection map and boundary quality assessment report, multi-layer wavelet decomposition is performed on the precisely registered image pair, and the image is decomposed into low-frequency approximation coefficients and high-frequency detail coefficients. For areas with high boundary quality (quality score > 0.8), a detail-preserving fusion strategy is adopted: the components with larger absolute values ​​of high-frequency coefficients are selected. For areas with low boundary quality, a spectrum-preserving fusion strategy is adopted: the low-frequency coefficients are weighted averaged to generate a multi-scale wavelet coefficient set.

[0156] S413: Design a boundary quality-aware fusion weight allocation algorithm, read the boundary positioning accuracy information Accuracy_boundary in the boundary quality assessment report, and calculate the local fusion weight w_local = Accuracy_boundary × w_base + (1-Accuracy_boundary) × w smooth , where w_base holds the boundary information, w smooth Smoothing is performed and dynamic weights are applied to the multi-scale wavelet coefficient set for coefficient fusion, which avoids the boundary blurring problem caused by traditional fixed weights and generates quality-aware fusion coefficients.

[0157] S414: Perform inverse wavelet transform on the quality-aware fusion coefficients to reconstruct the image. In the reconstruction process, boundary preservation constraints are introduced. Through iterative optimization, the reconstructed image is ensured to maximize boundary clarity while preserving spectral information. The spectral fidelity RMSE_spectral and spatial detail preservation SSIM of the fused image are calculated. spatial , verify that the fusion effect meets the quality requirements and output the fusion situation image.

[0158] By establishing a boundary-sensitive adaptive wavelet basis selection mechanism and implementing quality-aware fusion weight allocation in combination with boundary quality assessment reports, the traditional wavelet fusion method has fundamentally improved the technical defects of ringing effect and boundary blur in sharply changing boundary areas. This innovation uses Daubechies wavelet to maintain spectral fidelity in smooth areas and Biorthogonal wavelet to reduce ringing effect in sharply changing areas. At the same time, the local fusion weight w_local =Accuracy_boundary × w_base + (1-Accuracy_boundary) × w smooth Achieve dynamic weight allocation for quality perception. This adaptive fusion strategy enables the system to select the optimal fusion parameters based on the boundary quality characteristics of different regions, maximizing boundary clarity while maintaining the authenticity of spectral information. In the fusion imaging of building damage at the disaster site, this method improved the spectral fidelity RMSE from 18.7 of the traditional method to 12.3, an improvement of 34%, and the spatial detail preservation SSIM from 0.763 to 0.892, an improvement of 17%. This effectively solves the technical contradiction that the traditional fixed parameter fusion method cannot take into account both spectral fidelity and spatial details, and provides a high-quality fusion image foundation for accurate quantitative analysis of disaster losses.

[0159] According to one aspect of the present application, S42: Quantitative analysis of disaster losses, specifically:

[0160] S421: Read the fused situation image and boundary quality assessment report, and identify the areas of ground feature changes in the image. The traditional loss calculation based on pixel statistics ignores the impact of boundary accuracy on area calculation. Therefore, a boundary quality-aware area calculation method is introduced: for areas with high boundary quality (Accuracy>0.9), precise pixel counting is used. For areas with medium boundary quality (0.7≤Accuracy≤0.9), the uncertainty coefficient Area_adjusted = Area_pixel ×(0.85 + 0.15×Accuracy) is introduced to generate a quality-aware change area map.

[0161] S422: Based on the quality perception change area map, different types of disaster loss indicators are classified and calculated: the area of ​​building collapse is calculated by identifying the complete building boundary through connected domain analysis, the length of road blockage is calculated by the central axis length through the skeleton extraction algorithm, and the range of vegetation damage is calculated through spectral difference analysis combined with boundary quality weighting. Each loss indicator is associated with a corresponding boundary quality score for reliability assessment, and a classified loss statistics table is established.

[0162] S423: Based on the classified loss statistics table and boundary quality information, a loss degree grading evaluation system is established: slight loss (quality score > 0.8 and loss area < 10% of the total area), moderate loss (quality score > 0.6 and loss area 10%-30%), and severe loss (quality score > 0.4 and loss area > 30%). Areas with low quality scores are marked as "uncertain loss" and require further verification. The area, proportion and confidence level of each level of loss are calculated, and the disaster loss assessment results are output.

[0163] By introducing a boundary-quality-aware area calculation method (Area_adjusted = Area_pixel × (0.85 + 0.15 × Accuracy)) and establishing a loss degree grading system that includes confidence assessment, this innovatively addresses the key issue of traditional pixel-based loss calculations, which ignore the impact of boundary accuracy and lead to unreliable assessment results. This technology transforms abstract loss assessments into concrete quality-aware calculations, using precise pixel counting for areas with high boundary quality and introducing uncertainty coefficients for area adjustment in areas with medium boundary quality. Each loss indicator is associated with a corresponding boundary quality score for reliability assessment. The grading evaluation system categorizes the degree of loss into four levels: mild, moderate, severe, and uncertain. It calculates the area, proportion, and confidence level for each level, providing quantitative credibility indicators for the assessment results. In earthquake building damage assessment, this method reduces the accuracy of loss area calculation from the traditional method's ±25% error to ±8%, an improvement of 68%. At the same time, it provides a confidence interval of 0.6-0.92 for each assessment result, completely solving the technical defect of the traditional loss assessment method's lack of reliability quantification, and providing a statistically significant quantitative assessment basis for disaster relief decision-making and insurance claims.

[0164] In another embodiment of the present application, steps S3 and S4 can also be implemented by using spatiotemporal evidence fusion.

[0165] A method for collaboratively processing multi-dimensional images of disaster sites based on UAV-satellite collaboration includes the following steps:

[0166] Acquire multi-source heterogeneous image data, perform correction and communication delay compensation, establish a unified spatiotemporal reference system, and generate a spatiotemporal corrected image dataset;

[0167] Based on a spatiotemporal corrected image dataset, a multidimensional evidence fusion model is constructed through spatiotemporal evidence field modeling. An adaptive kernel selection mechanism is used to implement evidence-driven boundary reconstruction. The optimal boundary position is iteratively solved under a constrained optimization framework to generate an optimized reconstructed boundary map and a boundary reconstruction confidence map.

[0168] Confidence-weighted multidimensional image fusion is performed based on the optimized reconstructed boundary map and the boundary reconstruction confidence map, and a confidence-weighted loss quantitative calculation model is established to output quality-aware fused images and disaster loss assessment results.

[0169] Furthermore, the multi-dimensional evidence fusion model is constructed through spatiotemporal evidence field modeling, including:

[0170] Based on the time difference parameters in the spatiotemporal correction image dataset, the nonlinear evolution characteristics of time evidence are described by the disaster dynamics function, and the time evidence field is established in combination with the local change rate.

[0171] Calculate the structure tensor of each pixel in the spatiotemporally corrected image dataset and establish a spatial evidence field through anisotropy and coherence metrics;

[0172] The motion vector field is calculated by optical flow estimation, and the temporal and spatial coherence metrics are established, which are then fused to generate a dynamic evidence field.

[0173] A multi-scale evidence fusion operator is designed to fuse the temporal evidence field, spatial evidence field and dynamic evidence field at multiple scale levels to generate a multi-scale fused evidence field.

[0174] Preferably, implementing evidence-driven boundary reconstruction using an adaptive kernel selection mechanism includes:

[0175] Based on the multi-scale fusion evidence field, a threshold hierarchical system for evidence strength is established. The preservation kernel, bilateral filter kernel, or Gaussian kernel is adaptively selected according to the evidence strength interval. The anisotropic weight matrix is ​​calculated based on the local gradient direction to generate an adaptive reconstruction kernel set.

[0176] Combining the multi-scale fusion evidence field and the adaptive reconstruction kernel set, the basic confidence is calculated by the local structure complexity and boundary clarity, and the spatiotemporal consistency correction factor is introduced to generate the adaptive reconstruction weight matrix.

[0177] The improved Canny edge detection algorithm is used to extract the initial boundary of the spatiotemporal correction image dataset, and a dual-threshold adaptive adjustment mechanism is designed to generate the initial boundary estimation map.

[0178] Preferably, iteratively solving the optimal boundary position under the constrained optimization framework includes:

[0179] Based on the adaptive reconstruction weight matrix, a multi-objective constrained optimization objective function is designed, which includes fidelity term, smoothness term, time consistency term and structure preservation term, and a multi-objective optimization model is established;

[0180] Adopting a multi-objective optimization model to design an adaptive gradient descent iterative solution algorithm, the objective function gradient is calculated by the variational method, an adaptive step-size strategy is designed, and a constrained projection operator is implemented to ensure spatial continuity and temporal consistency.

[0181] Starting from the initial boundary estimation map, the boundary position is iteratively updated until the convergence condition is met, and the optimal reconstructed boundary map is output.

[0182] Preferably, generating the optimized reconstructed boundary map and the boundary reconstruction confidence map includes:

[0183] A multidimensional quality assessment system is established based on the optimal reconstructed boundary map and the spatiotemporal correction image dataset to calculate the reconstruction fidelity, boundary clarity, spatiotemporal consistency and structure preservation, and generate a reconstruction quality assessment vector.

[0184] An uncertainty quantification model is designed based on the reconstruction quality assessment vector. The confidence interval of each boundary point is calculated through Bayesian reasoning. A confidence propagation network is established to transfer the confidence of the spatial neighborhood and generate a boundary reconstruction confidence map.

[0185] A quality-driven parameter adaptive update mechanism is designed based on the boundary reconstruction confidence map and the reconstruction quality evaluation vector. When the quality score is lower than the threshold, parameter reoptimization is triggered and the iterative reconstruction process is re-executed to output the optimized reconstructed boundary map.

[0186] Preferably, performing confidence-weighted multi-dimensional image fusion based on the optimized reconstructed boundary map and the boundary reconstruction confidence map includes:

[0187] The optimal wavelet basis is selected based on the boundary reconstruction confidence map. Daubechies wavelet is used to preserve details in high-confidence areas, Biorthogonal wavelet is used to reduce ringing in medium-confidence areas, and Coiflets wavelet is used to smooth low-confidence areas. The quality-aware fused image is generated by combining the spatiotemporally corrected image dataset.

[0188] Based on the quality-aware fusion image, a confidence-weighted loss quantitative calculation model is established. The confidence-weighted area of ​​the building change area, the confidence-weighted length of the road blockage, and the confidence-weighted range of the vegetation damage are calculated. A classified loss statistical table is established and associated with the confidence intervals to output the disaster loss assessment results.

[0189] In another embodiment of the present application, specifically:

[0190] S3: Adaptive boundary reconstruction based on spatiotemporal evidence fusion

[0191] S31: Modeling and Initialization of Multidimensional Spatiotemporal Evidence Fields

[0192] S311: Read the spatiotemporal correction image dataset and time difference parameters, build the time evolution driven evidence field model, and use the disaster dynamics function T_dynamics(Δt) = (Δt / T_reference)^α × exp(-β×Δt2 ) describes the nonlinear evolution characteristics of temporal evidence, where α controls the evolution rate and β controls the time attenuation. Combined with the local change rate local_change_rate(x,y) = ||∇I2(x,y) - ∇I1(x,y)||, the temporal evidence field T_field(x,y) is established.

[0193] S312: Based on the spatiotemporal correction image dataset, a structure tensor driven spatial evidence field is constructed to calculate the structure tensor ST(x,y) = [I_x 2 I_x×I_y; I_x×I_y I_y 2 ], through the anisotropy metric anisotropy(x,y) = (λ1-λ2) / (λ1+λ2) and the coherence metric coherence(x,y) = (λ1-λ2) 2 / (λ1+λ2) 2 A spatial evidence field S_field(x,y) reflecting the boundary structure characteristics is established, where λ1 and λ2 are the eigenvalues ​​of the structure tensor.

[0194] S313: Combined with the sequence information of the spatiotemporally corrected image dataset, a motion consistency-driven dynamic evidence field is constructed. The motion vector field motion_vector(x,y) is calculated through optical flow estimation. The temporal coherence metric temporal_coherence = correlation(motion_t1, motion_t2) and the spatial coherence metric spatial_coherence = smoothness(motion_field) are established. The fusion generates a dynamic evidence field D_field(x,y) that describes the dynamic change pattern of the boundary.

[0195] S32: Adaptive Reconstruction Kernel Generation and Boundary Detection

[0196] S321: Read the temporal evidence field T_field, the spatial evidence field S_field, and the dynamic evidence field D_field, design a multi-scale evidence fusion operator, decompose each evidence field into multiple scale levels through scale space decomposition, and perform evidence fusion E_multi^k(x,y) = w_T^k×T_field^k + w_S^k×S_field^k + w_D^k×D_field^k at each scale k, where the weights w_i^k are adaptively adjusted according to the scale characteristics, and the multi-scale fused evidence field E_multi(x,y) is reconstructed across scales.

[0197] S322: Based on the multi-scale fusion evidence field E_multi, an adaptive kernel selection mechanism is designed to establish a hierarchical system of evidence strength thresholds: in high-evidence areas (E_multi > τ_high), the preservation kernel K_preservation is selected to preserve true changes; in medium-evidence areas (τ_low ≤ E_multi ≤ τ_high), the bilateral filter kernel K_bilateral is selected to achieve boundary smoothing; in low-evidence areas (E_multi < τ_low), the Gaussian kernel K_gaussian is selected for noise suppression. At the same time, the anisotropic weight matrix W_anisotropic(θ_local) is calculated according to the local gradient direction θ_local to generate the adaptive reconstruction kernel set K_adaptive.

[0198] S323: Combining the multi-scale fusion evidence field E_multi and the adaptive reconstruction kernel set K_adaptive, a confidence-driven reconstruction weight calculation model is established. The basic confidence base_confidence is calculated through the local structure complexity complexity_local = entropy(local_neighborhood) and the boundary sharpness_local = max(gradient_magnitude). The spatiotemporal consistency correction factor consistency_factor = correlation(temporal_neighbor) × smoothness(spatial_neighbor) is introduced to generate the adaptive reconstruction weight matrix W_recon(x,y).

[0199] S33: Iterative Boundary Reconstruction in a Constrained Optimization Framework

[0200] S331: Read the spatiotemporally corrected image dataset, perform initial boundary extraction using the improved Canny edge detection algorithm, and design a dual-threshold adaptive adjustment mechanism: the high threshold T_high = μ_gradient + 2×σ_gradient, the low threshold T_low = 0.4×T_high, where μ_gradient and σ_gradient are the mean and standard deviation of the gradient amplitude. Combined with non-maximum suppression and hysteresis threshold connection, generate the initial boundary estimation map B0(x,y).

[0201] S332: Based on the adaptive reconstruction weight matrix W_recon, design the multi-objective constrained optimization objective function E_objective = λ1×E_fidelity + λ2×E_smoothness + λ3×E_temporal + λ4×E_structure, where the fidelity term E_fidelity = ||B - B_observed|| 2 _W measures the consistency between the reconstructed boundary and the observed boundary, and the smoothness term E_smoothness = ∫||∇ 2 B|| 2 dx dy ensures boundary continuity, and the temporal consistency term E_temporal = ||B_t2 - T(B_t1)|| 2 Measuring time domain coherence, the structure preservation term E_structure = ||ST(B)- ST(B_reference)|| 2 _F Maintain local structural characteristics and establish a multi-objective optimization model.

[0202] S333: Adopt a multi-objective optimization model and design an adaptive gradient descent iterative solution algorithm. Calculate the objective function gradient ∇E_objective through the variational method. Design an adaptive step size strategy α_k = α0 / (1 + γ×k)×quality_feedback_factor, where quality_feedback_factor is dynamically adjusted according to the reconstruction quality. Execute the constraint projection operator P_constraints to ensure spatial continuity constraint_spatial: ||∇B|| < δ_smooth and temporal consistency constraint_temporal: correlation(B_k, B_{k-1}) > τ_consistency. Iteratively update B_{k+1} = P_constraints(B_k - α_k×∇E_objective(B_k)) until the convergence condition ||B_{k+1} -B_k||_∞ < ε is met. Output the optimal reconstructed boundary map B_final.

[0203] S4: Quality-aware fusion and adaptive evaluation output

[0204] S41: Reconstruction Quality Self-Assessment and Confidence Calculation

[0205] S411: Read the optimal reconstructed boundary map B_final and the spatiotemporal correction image dataset, establish a multi-dimensional quality assessment system, and calculate the reconstruction fidelity fidelity_score = 1 - ||B_final - B_ground_truth|| 2 / ||B_ground_truth|| 2 , boundary clarity sharpness_score = mean(gradient_magnitude(B_final)), spatiotemporal consistency consistency_score = 0.6×temporal_correlation + 0.4×spatial_smoothness, structure preservation structure_score = SSIM(structure_tensor(B_final), structure_tensor(B_reference)), generate the reconstruction quality evaluation vector Q_recon.

[0206] S412: Based on the reconstruction quality assessment vector Q_recon, an uncertainty quantification model is designed. The confidence interval of each boundary point is calculated through Bayesian reasoning. A confidence propagation network propagation_network is established for spatial neighborhood confidence transfer. The global confidence global_confidence = weighted_average(local_confidence, spatial_weights) is calculated. At the same time, the temporal domain confidence correction temporal_confidence_correction = correlation(confidence_t1, confidence_t2) is introduced to generate the boundary reconstruction confidence map C_boundary(x,y).

[0207] S42: Adaptive Parameter Update and Quality Feedback

[0208] S421: Read the boundary reconstruction confidence map C_boundary and the reconstruction quality evaluation vector Q_recon, design a quality-driven parameter adaptive update mechanism, increase the fidelity weight λ1 = λ1 × (1 + α_fidelity×quality_gap) when the reconstruction fidelity_score < threshold_fidelity, adjust the kernel parameter kernel_size = kernel_size ×adaptation_factor when the boundary sharpness_score < threshold_sharpness, correct the spatiotemporal weight ratio when the consistency_score < threshold_consistency, and establish an adaptive parameter update strategy.

[0209] S422: Based on the adaptive parameter update strategy, a quality feedback loop mechanism is established. When the overall quality score overall_quality = weighted_sum(Q_recon) < quality_threshold, parameter re-optimization is triggered and the iterative reconstruction process of S332-S333 is re-executed. At the same time, the parameter adjustment history is recorded to form learning samples. The parameter selection strategy is optimized through reinforcement learning methods to ensure the adaptive performance of the algorithm in different scenarios. The optimized reconstructed boundary map B_optimized and the quality assurance report Quality_report are output.

[0210] S43: Multidimensional Image Fusion and Loss Assessment

[0211] S431: Read the optimized reconstructed boundary map B_optimized, the quality assurance report Quality_report, and the spatiotemporally corrected image dataset, design a boundary quality-aware adaptive wavelet fusion algorithm, and select the optimal wavelet basis based on the boundary reconstruction confidence map C_boundary: use Daubechies wavelet to preserve details in high confidence areas (C_boundary > 0.85), use Biorthogonal wavelet to reduce ringing in medium confidence areas (0.6 ≤ C_boundary ≤ 0.85), and use Coiflets wavelet smoothing in low confidence areas (C_boundary < 0.6) to generate the quality-aware fused image I_fused.

[0212] S432: Based on the quality-aware fused image I_fused and the quality assurance report Quality_report, a confidence-weighted loss quantitative calculation model is established. For building change areas, the confidence-weighted area (Area_building = ∑(pixel_area × C_boundary(x,y)) is calculated. Road blockages are extracted by skeleton extraction and the confidence-weighted length (Length_road = ∫C_boundary(s)ds) is calculated. For vegetation damage, spectral analysis combined with confidence is used to calculate Range_vegetation = ∑(spectral_change × C_boundary × vegetation_mask). A classification loss statistics table (Loss_statistics) is established and associated with corresponding confidence intervals (Confidence_intervals). The disaster loss assessment result (Assessment_result) and the processing quality verification report (Verification_report) are output.

[0213] This implementation case study assesses building damage following a magnitude 7.2 earthquake in a mountainous area. The earthquake occurs at 14:30:00 UTC on May 20, 2025, using T_disaster. Multi-source imagery collected 30 and 90 minutes after the earthquake is processed to assess building collapse.

[0214] S1: Multi-source heterogeneous data acquisition and synchronous preprocessing

[0215] S11-S12: Get raw image data

[0216] Satellite imagery: Sourced from the Sentinel-2 satellite, acquired at T_sat = 14:45:00 UTC, with a resolution of 10 meters and including four bands: RGB and near-infrared. Drone imagery: Sourced from five quadrotor drones equipped with RGB cameras, acquired at T_uav = 15:00:00 UTC, with a resolution of 0.1 meters. Overlap area: 1.2 square kilometers, with an overlap ratio of 75%.

[0217] S13: Communication link status parameter collection. Real-time monitoring of the network status parameters is as follows: Satellite link: Bandwidth occupancy BW_sat = 0.45, signal strength -82dBm, transmission delay τ_sat_measured = 580ms. UAV direct link: Bandwidth occupancy BW_uav = 0.32, signal strength -65dBm, transmission delay τ_uav_measured = 45ms. Ground base station link: Bandwidth occupancy BW_gnd = 0.67, signal strength -58dBm, transmission delay τ_gnd_measured = 120ms.

[0218] S14: Image normalization

[0219] The satellite image was geometrically corrected with the Brown distortion model parameters: k1 = -0.0012, k2 = 0.0000045. The radial distortion correction results showed that the edge area position offset was reduced by 85%. The drone image was radiometrically corrected with the dark current value DN_dark = 20 and the gain coefficient Gain = 0. factor = 0.125, and the corrected image dynamic range is increased to 0-4095.

[0220] S2: Heterogeneous Timestamp Correction and Communication Delay Compensation

[0221] S211: Specific calculation process of delay prediction model construction

[0222] We extracted the delay data for the past 24 hours from the historical communication link status parameter table and analyzed it by grouping it into 10-minute time windows:

[0223] Satellite link historical delay analysis: delay mean τ_sat mean = 565ms; standard deviation τ_sat std = 45ms; change trend slope τ _sat = 2.3ms / hour (increasing trend);

[0224] Historical delay analysis of drone links: average delay τ_uav mean = 38ms; standard deviation τ_uav std = 12ms; Slope τ _uav = -0.8ms / hour (decreasing trend);

[0225] S212: Delayed Impact Factor Model Establishment

[0226] The piecewise linear relationship model is constructed in step S212: When BW currentWhen < 0.6, τ is obtained by least squares fitting. BW = k1×BW current + b1; where k1 = 180ms, b1 = 45ms; when BW current ≥ 0.6, τ BW = k² × BW current + b2; where k2 = 420ms, b2 = -105ms; network load impact model τ Load = k3×Load current 2 + k4×Load current + b3; where k3 = 85ms, k4 = 125ms, b3 = 15ms; Load current is the current network load rate.

[0227] For the current satellite link (BW_sat = 0.45 < 0.6), calculate: τ BW _sat = 180 × 0.45 + 45 = 81 + 45 = 126ms.

[0228] Assuming the current network load Load_sat = 0.35, calculate: τ Load _sat = 85 × 0.35 2 + 125 × 0.35 + 15 = 85 × 0.1225 + 43.75 + 15 = 10.41 + 43.75 + 15 = 69.16ms.

[0229] S213: Comprehensive delay prediction function calculation

[0230] Comprehensive delay prediction function τ pred = α×τ hist + β×τ BW + γ×τ Load + δ×trend factor ; where α = 0.4, β = 0.3, γ = 0.2, δ = 0.1 are weight coefficients; τ hist is the historical delay mean; τ BW The bandwidth affects the delay; τ Load The load affects the delay; trend factor It is a trend correction item.

[0231] In this embodiment, the satellite link is calculated: trend factor _sat = slope τ_sat × 1 = 2.3 × 1 = 2.3ms (calculated based on 1 hour). τ pred _sat = 0.4 × 565 + 0.3 × 126 + 0.2 × 69.16 +0.1 × 2.3. τ pred _sat = 226 + 37.8 + 13.83 + 0.23 = 277.86ms.

[0232] S221: Timestamp backtracking correction calculation

[0233] Delay compensation value calculation: satellite image T_sat received = 14:45:30.580 UTC, delay compensation τ applied comp _sat = 277.86ms; corrected timestamp T_sat corrected = T_sat received - τ comp _sat = 14:45:30.580 - 0.27786 = 14:45:30.302 UTC; T_uav is calculated similarly for the drone image corrected = 15:00:28.156 UTC.

[0234] Verified by reference images, the mean absolute error (MAE) of timestamp correction is 28ms < 50ms threshold, which meets the subsequent registration accuracy requirements.

[0235] S3: Multidimensional Image Registration Driven by Boundary Evidence Strength

[0236] S31: Select the image pair with the largest time interval from the spatiotemporal correction image dataset: satellite image T_sat corrected = 14:45:30.302 and drone image T_uav corrected = 15:00:28.156, with a time difference of Δt = 894.854 seconds ≈ 14.91 minutes. Analysis of the overlapping areas shows an overlap rate of 75% and a feature density of 156 features per square kilometer.

[0237] S32: Boundary Evidence Strength Assessment Mechanism

[0238] S321: Perform Sobel edge detection on the image pair to be registered, with the threshold Th_grad set to 50: 2847 boundary points are extracted from the satellite image. 12635 boundary points are extracted from the drone image. There are 1923 common boundary points in the overlapping area.

[0239] We selected a typical boundary point P (x=256, y=128) for detailed calculations: the gradient magnitude at that point in the satellite image is: gradI_sat(256,128) = 78. The gradient magnitude at the corresponding point in the drone image is: gradI_uav(256,128) = 145. The difference in gradient direction angle is: Δθ = 32°.

[0240] S322: Temporal Evidence Strength Calculation

[0241] Temporal evidence strength S time = |Δt| / T threshold ; where Δt = 894.854 seconds is the time difference; T threshold is the time threshold parameter, and the earthquake scenario sets T threshold = 1800 seconds (30 minutes); S time = 894.854 / 1800 = 0.497.

[0242] S323: Spatial Evidence Strength Calculation

[0243] Spatial evidence strength S spatial = |gradI1(x,y) - gradI2(x,y)| / max(gradI1(x,y),gradI2(x,y)); where gradI1(256,128) = 78 is the gradient amplitude of the satellite image; gradI2(256,128) = 145 is the gradient amplitude of the drone image; S spatial = |78 - 145| / max(78, 145) = 67 / 145 = 0.462.

[0244] S324: Calculation of Consistency Evidence Strength

[0245] A 7×7 window is built around the boundary point P(256,128) to analyze the change pattern of 49 pixels; S consist =(N consist ent - N isolated ) / N total ; where N consist ent = 31 pixels are consistent with the neighborhood change direction; N isolated =5 isolated change pixels; N total = 49 total pixels; S consist = (31 - 5) / 49 = 26 / 49 = 0.531.

[0246] S325: Calculation of the overall strength of evidence

[0247] Overall strength of evidence Stotal = w1×S time + w2×S spatial + w3×S consist ; Among them, the weight parameters w1 = 0.4, w2 = 0.4, w3 = 0.2; S time = 0.497 is the time evidence strength; S spatial = 0.462 is the spatial evidence strength; S consist = 0.531 is the strength of evidence for consistency; S total = 0.4×0.497 + 0.4×0.462 + 0.2×0.531 =0.1988 + 0.1848 + 0.1062 = 0.490.

[0248] S33: Segment boundary processing strategy

[0249] Connected domain analysis identified 89 continuous boundary regions with a total length of 23784 pixels. Segmentation by 4-pixel length resulted in 5946 boundary segments. The boundary segment where the boundary point P(256,128) is located is assigned an identifier ID. seg = 1247.

[0250] S332: Boundary Segment Classification

[0251] Average strength of evidence for boundary segments: seg = 1247, the evidence strength of the four pixel points is 0.490, 0.523, 0.456, and 0.478 respectively; seg = (1 / N)∑S total (i) = (0.490 + 0.523 + 0.456 +0.478) / 4 = 1.947 / 4 = 0.487; since 0.3 ≤ S seg = 0.487 ≤ 0.7, the segment was classified as a moderate strength of evidence segment.

[0252] Statistics of all boundary segment classification results: high evidence strength segment (S seg > 0.7): 1245 segments, accounting for 20.9%. Moderate evidence strength segment (0.3 ≤ S seg ≤ 0.7): 3567 segments, accounting for 60.0%. Low evidence strength segment (S seg < 0.3): 1134 segments, accounting for 19.1%.

[0253] S333: Differentiated Processing Strategy Implementation

[0254] ID seg= 1247 medium-intensity evidence segment, using weighted fusion of evidence strength:

[0255] Moderate-strength evidence segment fusion I fused (x,y) = S seg ×I new (x,y) + (1-S seg )×I old (x,y); where S seg = 0.487 is the average strength of evidence for the segment; I new (256,128) = 156 is the pixel value of the drone image; I old (256,128) = 89 is the pixel value of the satellite image; I fused (256,128) = 0.487×156 + (1-0.487)×89 = 75.972 +45.657 = 121.629 ≈ 122.

[0256] S334: Spatial splicing of segment processing results

[0257] After the stitching is completed, 67 boundary jumps are detected, with the maximum jump amplitude being 32 grayscale values, which are recorded in the boundary jump record table.

[0258] S34: Boundary Segment Constraint Propagation

[0259] S341: Construction of inter-segment constraint relationship diagram

[0260] Adjacent boundary segment ID seg = 1247 and ID seg = 1248 for analysis: Segment 1247 is classified as a moderate-evidence segment (category code = 2). Segment 1248 is classified as a high-evidence segment (category code = 3). Treatment strategy difference D_diff = |category i - Category j | = |2 - 3| = 1; since D_diff = 1 < 2, the constraint adjustment mechanism is not triggered.

[0261] However, for segments 1247 and 1249: Segment 1249 classification: low evidence strength segment (category code = 1). D_diff = |2 - 1| = 1 < 2, and no adjustment is triggered.

[0262] For segments 1250 and 1251: Segment 1250: High-evidence segment (category = 3). Segment 1251: Low-evidence segment (category = 1). D_diff = |3 - 1| = 2, triggering constraint adjustment.

[0263] S342: Continuity penalty function calculation

[0264] Continuity penalty function P penalty = exp(D_diff×λ)×|I i -I j |; where D_diff = 2 is the strategy difference; λ = 0.5 is the penalty coefficient; I i = 178 is the pixel value of the high-evidence segment boundary; I j = 98 is the pixel value of the low-evidence segment boundary; P penalty = exp(2×0.5)×|178 - 98| = exp(1.0)×80 = 2.718×80 = 217.44; Set the threshold Th penalty = 150, due to P penalty = 217.44 > 150, triggering the constraint adjustment mechanism.

[0265] S343: Constraint adjustment instruction execution

[0266] Segment 1250 is adjusted from high evidence strength to medium evidence strength, and segment 1251 is adjusted from low evidence strength to medium evidence strength, and the fusion process is performed again.

[0267] S35: Sliding Window Boundary Reconstruction

[0268] S351: Sliding Window Settings

[0269] A sliding window of 9 pixels in length and 3 pixels in step size is created. Window W_1247, which contains boundary segments 1247-1249, is analyzed. The window contains 12 boundary points.

[0270] S352: Establishment of constraint optimization objective function

[0271] Constrained optimization objective function F opt = α × E evidence + β×E smooth + γ×E preserve ; where α = 0.5, β = 0.3, γ = 0.2 are weight parameters; E evidence Evidence retention item; smooth is the smoothing term; E preserve Keep the original information. Calculate the following items for window W_1247:

[0272] Evidence Preservation Item E evidence = ∑(S total (i) × |I new (i) - I current(i)|) = 0.487×|156-122| + 0.523×|167-134| + ... = 234.5. Smoothing term E smooth = ∑|I(i+1) - I(i)| = |122-134| + |134-128| + ... = 89.3. Original retained item E preserve = ∑|I current (i) - I_original(i)| = 45.7. Objective function value F opt = 0.5×234.5 + 0.3×89.3 + 0.2×45.7 = 117.25 + 26.79+ 9.14 = 153.18.

[0273] S353: Gradient Descent Optimization Solution

[0274] Use gradient descent optimization with a learning rate of η = 0.01 and a maximum number of iterations of 1000:

[0275] Iteration 1: gradF opt = [2.34, -1.67, 0.89, ...], Update I new = I old - η×gradF opt Iteration 15: ΔF = |F new -F old | = 0.0008 < ε_converge = 0.001, the algorithm converges.

[0276] The pixel values ​​of the boundary points after optimization are: [125, 131, 127, 135, 129, 138, 132, 141, 136].

[0277] S354: Overlapping area fusion processing

[0278] Calculate the weighted average of the distance: for the overlapping boundary point (258,128), affected by both windows W_1247 and W_1248; I final (258,128) = Σ(w i ×I i (258,128)) / Σw i ;where w_1247 = 1 / d_1247 = 1 / 1.41 =0.707, w_1248 = 1 / d_1248 = 1 / 2.83 = 0.354; I_1247(258,128) = 135, I_1248(258,128) = 142; Ifinal (258,128) = (0.707×135 + 0.354×142) / (0.707 + 0.354) =(95.445 + 50.268) / 1.061 = 145.713 / 1.061 = 137.3 ≈ 137.

[0279] S4: Collaborative integration and result output

[0280] S41: Multi-scale Image Fusion

[0281] S411: Adaptive wavelet basis selection

[0282] The sharp change boundary areas were detected to account for 23.5% of the total area. Biorthogonal wavelet was used in these areas, and Daubechies wavelet was used in the remaining areas.

[0283] S412-S414: Quality-aware fusion processing

[0284] Calculate the local fusion weight: for the area with boundary quality score Accuracy_boundary = 0.85; w_local = Accuracy_boundary × w_base + (1-Accuracy_boundary) × w smooth ; where w_base = 0.8, w smooth = 0.3; w_local = 0.85×0.8 + (1-0.85)×0.3 = 0.68 + 0.045 = 0.725.

[0285] Fusion result quality assessment: spectral fidelity RMSE_spectral = 12.3 (better than the traditional method's 18.7). Spatial detail preservation SSIM spatial = 0.892 (better than 0.763 of the traditional method).

[0286] S42: Quantitative Analysis of Disaster Losses

[0287] S421: Quality-Aware Area Calculation

[0288] For the building change area with boundary quality Accuracy = 0.85:

[0289] Calculate the adjusted area: Since 0.7 ≤ Accuracy = 0.85 ≤ 0.9, it is a medium-quality area; Area_adjusted = Area_pixel × (0.85 + 0.15×Accuracy) = 2847×(0.85 + 0.15×0.85) = 2847×(0.85 + 0.1275) = 2847×0.9775 = 2782.4 square meters.

[0290] S422-S423: Classification loss statistics

[0291] Loss statistics: Building collapse area: 2,782.4 square meters (confidence level 0.85). Road blocked length: 1.2 kilometers (confidence level 0.78). Vegetation damaged area: 0.8 square kilometers (confidence level 0.92).

[0292] Grading assessment results: Severe damage area: 45.2% (confidence level > 0.8); Moderate damage area: 32.7% (confidence level 0.6-0.8); Slight damage area: 18.3% (confidence level > 0.8); Uncertain damage area: 3.8% (confidence level < 0.6).

[0293] Compared with the traditional SIFT registration method, registration accuracy improved from 3.2 pixels to 0.8 pixels, a 75% improvement. Boundary continuity: The change in boundary curvature was reduced by 60%, eliminating the "ghost boundary" phenomenon. Loss assessment accuracy increased from 67% to 89.3%, a 33.2% improvement.

[0294] In another embodiment of the present application, Sentinel-1 SAR imagery was received via a portable satellite communication terminal, KaSat-P300, at an acquisition time of T_sat = 03:00:00 UTC, with a resolution of 5 meters and C-band VV and VH polarization data. UAV imagery was also transmitted via a mesh network mode using five hexacopter drones equipped with dual RGB and thermal infrared sensors, at an acquisition time of T_uav = 03:30:00 UTC, with a visible light resolution of 0.08 meters and a thermal infrared resolution of 0.32 meters.

[0295] Portable satellite terminal KaSat-P300 link parameters: Ka-band uplink bandwidth occupancy BW_uplink = 0.72, downlink bandwidth occupancy BW_downlink = 0.58, signal strength RSSI = -78dBm, signal attenuation due to typhoon attenuation_weather = 8.5dB, transmission delay τ_sat_portable = 620ms.

[0296] Portable terminal MESH network parameters: A self-organizing network built by five drone nodes through portable terminals. The network topology is a star structure, with the central node being the ground portable terminal GT-Mesh-500. The distance between each drone node is 150-300 meters, the hop count is 1-2 hops, the network load is load_mesh = 0.45, and the average transmission delay is τ_mesh = 85ms.

[0297] Backup BGAN link parameters: The Hughes 9502 portable BGAN terminal is used as the backup link, operating in the L-band, with a bandwidth of 384 kbps, a signal strength of -82 dBm, and a transmission delay of τ_bgan = 1200 ms.

[0298] For portable Ka-band satellite terminals, considering the impact of signal attenuation in typhoon environments, a weather correction delay model is established: τ_sat_weather = τ_base_sat × (1 + weather_attenuation_factor), where weather_attenuation_factor = 0.15 × (attenuation_weather / 10dB). The calculated delay is τ_sat_weather = 565 × (1 + 0.15 × 8.5 / 10) = 565 × 1.1275 = 637ms.

[0299] For portable terminal MESH networks, a multi-hop delay accumulation model is established: τ_mesh_total = τ_processing × hop_count + τ_transmission × link_quality_factor, where τ_processing = 15ms is the per-hop processing delay, and link_quality_factor is dynamically adjusted based on signal strength. The calculated value is τ_mesh_total = 15 × 2 + 55 × 0.92 = 30 + 50.6 = 80.6ms.

[0300] Multi-link delay prediction for portable terminals: The predicted delay for the primary link (Ka satellite) is τ_pred_ka = 637ms, the predicted delay for the backup link (BGAN) is τ_pred_bgan = 1200ms, and the predicted delay for the MESH network is τ_pred_mesh = 80.6ms.

[0301] Timestamp correction calculation: Satellite SAR image correction T_sat_corrected = 03:00:42.637 - 0.637 = 03:00:42.000 UTC, UAV image correction T_uav_corrected = 03:30:28.081 - 0.081 = 03:30:28.000 UTC, time difference Δt = 1800 seconds = 30 minutes.

[0302] The processing flow is the same as that of S31-S33 in the above embodiment, but the algorithm is adapted to the scattering characteristics of the SAR image:

[0303] Temporal evidence field calculation: Considering the changes in ground scattering characteristics in a typhoon environment, T_field_sar(x,y) = (Δt / T_threshold_typhoon) × scatter_coherence_factor, where T_threshold_typhoon = 2400 seconds (40 minutes) and scatter_coherence_factor reflects the SAR scattering consistency.

[0304] Spatial evidence field adaptation: For the coherent speckle noise of SAR images, Lee filtering is used for preprocessing and then the structure tensor is calculated, S_field_sar(x,y) = anisotropy_lee_filtered × gradient_magnitude_ratio_sar.

[0305] The strength of evidence for the typical boundary point P_sar (x=312, y=156) is calculated as follows: S_time = 1800 / 2400 = 0.75, S_spatial_sar = 0.523, S_consist = 0.612, and the comprehensive strength of evidence S_total_sar = 0.4×0.75 + 0.4×0.523 + 0.2×0.612 = 0.631.

[0306] Based on the limited bandwidth constraints of portable terminals, a compressed sensing fusion algorithm was designed: using wavelet-domain sparse representation, the fused image is compressed to 35% of the original data volume for transmission. At the receiver, compressed sensing is used to reconstruct the complete fused image. Taking into account the bandwidth limitations of portable terminals, a hierarchical transmission strategy was established: critical damage areas (collapsed buildings) are transmitted at their original resolution, general damage areas (vegetation damage) are transmitted at 50% resolution, and intact areas are transmitted at 20% resolution.

[0307] Loss assessment results: The damaged area of ​​buildings is 3,247 square meters (confidence level 0.78), and the blocked road length is 2.1 kilometers (confidence level 0.82).

[0308] Effective in the harsh communication environments of disaster sites, through multi-link aggregation, adaptive compression, and a hierarchical transmission strategy, efficient data transmission and real-time loss assessment were achieved while ensuring image processing accuracy. Communication stability: Portable terminals maintained 89% link availability in typhoon conditions, while ground base stations were completely disrupted. Through multi-link aggregation and adaptive compression, data transmission efficiency was improved by 45%. Portable terminals could be deployed within two hours of a disaster, establishing an emergency communication network. Under bandwidth-constrained conditions, boundary reconstruction accuracy reached 1.2 pixels, a 33% reduction compared to the uncompressed optimization method, while increasing transmission efficiency threefold.

[0309] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the scope of protection of the present invention.

Claims

1. A collaborative multi-dimensional image processing method for disaster sites based on UAV-satellite collaboration, characterized by: The following steps are involved: Acquire multi-source heterogeneous image data, perform correction and communication delay compensation, establish a unified spatiotemporal reference system, and generate a spatiotemporal corrected image dataset; Based on a spatiotemporally corrected image dataset, a boundary evidence strength synthesis model is applied to decompose image boundaries into processing segments and a differentiated registration strategy is implemented based on the evidence strength. This eliminates discontinuous areas in the segmented processing and generates accurately registered image pairs and boundary quality assessment reports. Based on this, multi-dimensional image fusion processing is performed to generate a fused situation image, quantitatively analyze the loss area, and obtain the fused situation image and disaster loss assessment results.

2. The method according to claim 1, characterized in that A differentiated registration strategy was implemented using a comprehensive model of boundary evidence strength, including: Select the image pair with the largest time interval from the spatiotemporal correction image dataset as the image pair to be registered; A comprehensive boundary evidence strength model is constructed and used to evaluate the image pairs to be registered and generate a boundary evidence strength map. The continuous boundary is decomposed into processing segments based on the boundary evidence strength map, and differential fusion is implemented to generate segmented processing results.

3. The method according to claim 2, characterized in that Apply the boundary evidence strength comprehensive model to evaluate the image pairs to be registered and generate a boundary evidence strength map, including: Perform edge detection on the image pair to be registered, extract boundary pixels with gradient amplitude greater than the threshold, and establish a boundary pixel set and gradient feature record table; Calculate the temporal evidence strength of each boundary point in the boundary pixel point set based on the temporal difference between images and generate a temporal evidence strength graph; Calculate the gradient difference of each boundary point in the two images according to the gradient feature record table to generate a spatial evidence strength map; A window is established around each boundary point to analyze the spatial distribution pattern of boundary changes, calculate the consistency of neighborhood changes, and generate a consistency evidence strength map; The time, space and consistency evidence strength maps are weightedly fused to form a boundary evidence strength map.

4. The method according to claim 2, characterized in that Decompose continuous boundaries into processing segments based on boundary evidence strength maps and implement differential fusion, including: Perform connected domain analysis on the boundary evidence strength map to identify continuous boundary areas, divide them into equal-length segments, and generate a boundary segment identification table; Combined with the boundary segment identification table, the average evidence strength value of each boundary segment is calculated and compared with the threshold, and the boundary segments are divided into high evidence strength segments, medium evidence strength segments and low evidence strength segments to form a boundary segment classification table; For high evidence intensity segments, the pixel values ​​at the boundary of the new image are replaced. For medium evidence intensity segments, the evidence intensity weighted fusion is used. For low evidence intensity segments, Poisson gradient fusion is performed. The intermediate results of segmentation processing are generated and spatially spliced ​​to obtain the segmentation processing results.

5. The method according to claim 2, characterized in that Eliminate discontinuous areas in the segmentation process, generate accurately registered image pairs and boundary quality assessment reports, including: Through the constraint propagation network between boundary segments, the processing strategy differences of adjacent segments in the segmentation processing results are detected. When the difference exceeds the threshold, the processing method is adjusted to generate the constraint-adjusted boundary; The sliding window reconstruction technique is used to optimize the boundary after constraint adjustment, and boundary continuity is achieved under the constrained optimization framework to obtain accurately registered image pairs and boundary quality assessment reports.

6. The method according to claim 3, characterized in that Calculation of temporal, spatial, and consistency strength of evidence maps, including: Perform edge detection on the image pair to be registered, extract boundary pixels with gradient magnitude greater than the threshold, calculate the gradient direction angle of each boundary point, and establish a boundary pixel set and gradient feature record table; Based on the time difference between images, the time evidence strength of each boundary point in the boundary pixel point set is calculated as the ratio of the time difference to the time threshold, and a time evidence strength map is generated; Read the gradient feature record table, calculate the ratio of the gradient difference between the two images and the maximum gradient value of each boundary point as the spatial evidence strength, and generate a spatial evidence strength map; An n×n pixel window is established around each boundary point, and the ratio of the number of pixels in the window that are consistent with the neighborhood change direction to the total number of pixels is calculated as the consistency evidence strength to generate a consistency evidence strength map, where n is a preset value.

7. The method according to claim 2, characterized in that The process of constructing a comprehensive model of boundary evidence strength includes: Constructing a temporal evidence evaluation function based on the temporal differences in the image pair to be registered; Calculate image gradient differences and construct a spatial evidence evaluation function based on them; Analyze the coherence of local change patterns and construct a consistency evidence evaluation function based on it; Weights are configured for the temporal evidence evaluation function, spatial evidence evaluation function, and consistency evidence evaluation function, and weighted fusion is performed to form a comprehensive model of boundary evidence strength.

8. The method according to claim 5, characterized in that Generates constraint-adjusted bounds, including: Based on the segmentation processing results, an adjacency relationship graph between boundary segments is constructed, the difference in processing strategies between spatially adjacent boundary segments is calculated, and an inter-segment constraint relationship graph is generated; A continuity penalty function is designed based on the inter-segment constraint relationship graph. When the difference in processing strategies exceeds a threshold, a constraint adjustment mechanism is triggered and a constraint adjustment instruction table is generated. The segmentation processing results are locally reprocessed according to the constraint adjustment instruction table, the processing strategy of high-difference adjacent segments is adjusted to the intermediate category processing method, and the constraint-adjusted boundary is output.

9. The method according to claim 5, characterized in that Optimization using sliding window reconstruction technology, including: A sliding window with a length of a pixels is established for the boundary after constraint adjustment. The window step size is set to b pixels so that each boundary segment is completely covered, and a sliding window position table is generated; a and b are preset values; Based on the sliding window position table and the constraint-adjusted boundary, a constrained optimization objective function including evidence preservation term, smoothing term and original information preservation term is established for each sliding window to solve the optimal boundary position and generate a window optimization result set; The overlapping boundary areas in the window optimization result set are fused using the distance weighted average method to obtain accurately registered image pairs and boundary quality assessment reports.

10. The method according to claim 1, characterized in that Communication delay compensation, including: Build a delay prediction model based on the network status parameters of the communication link and calculate the delay compensation value of each transmission link; Apply the delay compensation value to retroactively correct the image transmission timestamp information to eliminate the systematic deviation of the timestamp caused by the delay difference of different communication links, and generate a correction timestamp record table; Based on the correction timestamp record table, a unified time axis based on the earliest image time is established, the image sequence is reorganized, and the spatiotemporal correction image dataset is output.

Images