Disaster site satellite communication blind area image acquisition and transmission method using low-altitude unmanned aerial vehicle relay

Abstract

The application provides a low-altitude unmanned aerial vehicle relay disaster site satellite communication blind area image acquisition and transmission method, comprising: obtaining disaster area information and performing adaptive multi-modal image acquisition; establishing a low-altitude unmanned aerial vehicle relay network; performing feature extraction and semantic analysis on the marked image dataset, generating non-uniform image patches and calculating transmission priority; monitoring the communication environment state, and dynamically adjusting the transmission strategy through a double-loop feedback mechanism. The application ensures semantic integrity, realizes transmission window awareness, stabilizes parameter adjustment through a double-loop feedback mechanism, and solves the problem of efficient transmission of edge images in the disaster site satellite communication blind area.

Description

Technical Field

[0001] This invention belongs to the field of satellite communication, and in particular to a method for acquiring and transmitting images of satellite communication blind spots at disaster sites via low-altitude unmanned aerial vehicle relay. Background Technology

[0002] In disaster relief, timely access to on-site information is crucial for rescue decision-making. Due to terrain obstructions and disaster damage, disaster areas often become satellite communication dead zones, making it difficult to transmit critical information in real time. Low-altitude drones, with their maneuverability and flexibility, can enter areas inaccessible to personnel to collect images and transmit information to communication coverage areas through relay networks, providing valuable visual data support for rescue command and significantly improving rescue efficiency and safety.

[0003] Currently, the application of drones in disaster relief has a certain research foundation. Traditional methods typically employ fixed flight path planning, combined with simple wireless relay modes for data transmission. Image acquisition often uses timed shooting with preset parameters, while data transmission generally uses uniformly segmented or progressively encoded transmission strategies. In terms of network construction, common drone relay systems are mostly based on static topologies or simple self-organizing protocols, adjusting their positions based on signal strength assessment. Regarding changes in communication quality, existing systems mostly employ reactive adjustments, maintaining communication based on reconnection mechanisms after link breaks.

[0004] However, in complex and ever-changing disaster environments, existing technologies face several insurmountable challenges: In the unique environment of the edge of satellite communication blind zones, communication quality exhibits non-linear fluctuations. Conventional single-cycle feedback control systems struggle to simultaneously meet the dual requirements of rapid parameter response and stable strategy optimization, leading to either slow system response resulting in missed transmission opportunities or frequent parameter oscillations affecting transmission stability. Simultaneously, traditional priority scheduling models are primarily designed based on content value and resource consumption, lacking the ability to perceive and model the time dimension of the transmission window. This prevents dynamic adjustment of transmission strategies at different stages of the window, particularly in the critical state before the communication window closes, hindering the priority transmission of high-value, small-volume content. Furthermore, existing image segmentation methods often rely on fixed grids or simple content features, failing to consider both semantic integrity and transmission environment adaptability in disaster scenarios. This makes it difficult to address the extremely uneven distribution of information value in disaster images, resulting in the segmentation of critical semantic information or the formation of isolated structures unfavorable for transmission. These intricate and critical technical issues restrict the efficient acquisition and transmission of disaster scene image information, urgently requiring innovative solutions. Summary of the Invention

[0005] The purpose of this invention is to provide a method for acquiring and transmitting images of disaster sites in satellite communication blind spots via low-altitude unmanned aerial vehicle (UAV) relay, in order to solve at least one technical problem existing in the prior art.

[0006] Technical solution: A method for image acquisition and transmission in satellite communication blind spots at disaster sites via low-altitude UAV relay, including:

[0007] Acquire disaster area information and perform adaptive multimodal image acquisition to generate a labeled image dataset;

[0008] A low-altitude UAV relay network was established based on a labeled image dataset, and the relay network topology was constructed and maintained.

[0009] Feature extraction and semantic analysis are performed on the labeled image dataset to generate non-uniform image segments, and the transmission priority of each segment is calculated to form a transmission priority sequence.

[0010] The system monitors the communication environment status of the relay network topology, combines the transmission priority sequence, and dynamically adjusts the transmission strategy through a double-loop feedback mechanism to perform adaptive transmission of image fragments.

[0011] Beneficial effects: This invention ensures that related content can be transmitted in a proper order by using non-uniform image segmentation, avoiding the problem of isolated information being difficult to understand; it also breaks through the limitations of traditional single-cycle feedback by using a dual-cycle feedback transmission scheduling mechanism, and realizes adaptive strategy switching during the transmission window stage. Attached Figure Description

[0012] Figure 1 A flowchart illustrating the steps of a method for acquiring and transmitting images of satellite communication blind spots at disaster sites via low-altitude UAV relay, as provided in this application embodiment.

[0013] Figure 2 A flowchart illustrating the steps of dynamically adjusting the transmission strategy through a dual-loop feedback mechanism, as provided in this application embodiment.

[0014] Figure 3 A flowchart illustrating the steps for generating a fine-grained scheduling plan provided in this application embodiment.

[0015] Figure 4 A flowchart illustrating the steps for generating transmission control instructions provided in this application embodiment. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] It should be noted that, for the purpose of clearly demonstrating the steps of this application, each step has been numbered in the specification. These numbers are for ease of explanation only and do not limit the execution order of the steps. In actual operation, depending on the technical requirements of the specific implementation scenario, the steps may be executed in a different order than that shown in the specification, and in some cases, parallel processing between steps can also be achieved.

[0018] like Figure 1 As shown, the method for acquiring and transmitting images of disaster sites in satellite communication blind spots via low-altitude UAV relay includes the following steps:

[0019] S1. Acquire disaster area information and perform adaptive multimodal image acquisition to generate a labeled image dataset;

[0020] Specifically, disaster area information includes geographical data, satellite coverage data, and meteorological environmental data. Geographical data can include the topography and location of the affected area; satellite coverage data uses satellite imagery to obtain an aerial view of the affected area, which can help understand the scale and impact of the disaster; meteorological environmental data can include weather conditions in the disaster area (such as wind speed and rainfall).

[0021] S2. Establish a low-altitude UAV relay network based on the labeled image dataset, and construct and maintain the relay network topology;

[0022] Specifically, drones will be deployed at low altitudes over disaster areas as signal relay points. They act as "flying signal towers," helping to overcome problems such as damaged ground-based signal towers or insufficient signal coverage. The network topology refers to the connection method between the various nodes (i.e., drones) in the drone relay network. Through an intelligent control system, drones can dynamically adjust their positions and connections to respond to changes in the environment or demand. This functions like an automatically adjusting communication network, ensuring signal coverage of the target area and efficient transmission.

[0023] S3. Perform feature extraction and semantic analysis on the labeled image dataset to generate non-uniform image segments, and calculate the transmission priority of each segment to form a transmission priority sequence.

[0024] Specifically, key feature information, such as the severity of the disaster and areas requiring priority, is extracted from labeled image datasets. Semantic analysis is used to understand the actual meaning of the image content, such as identifying which areas are emergency rescue points and which can be addressed later. Large images are divided into multiple smaller segments, each potentially containing varying degrees of information importance. Non-uniformity means that these segments differ in size or information density, depending on the importance or complexity of the region. The importance of each image segment is analyzed (e.g., which segments contain critical rescue areas), and priorities are assigned to each segment based on importance, determining which segments need to be transmitted first.

[0025] S4. Monitor the communication environment status of the relay network topology, combine the transmission priority sequence, and dynamically adjust the transmission strategy through a double-loop feedback mechanism to perform adaptive transmission of image fragments.

[0026] Specifically, the system continuously observes and analyzes the communication status of the UAV relay network, such as signal quality and connection stability. If network problems occur (e.g., signal interference or relay node failure), the system will promptly identify them. Utilizing a previously generated transmission priority sequence (ranking image fragments by importance), the system ensures that critical data is processed first during transmission; that is, more important information is scheduled for transmission first. A two-layer feedback mechanism is used to continuously adjust the strategy based on network communication status and transmission performance, making the transmission process both efficient and flexible.

[0027] This embodiment provides comprehensive and accurate information about disaster areas through adaptive multimodal image acquisition. A relay network established by low-altitude UAVs overcomes the problems of damage or insufficient coverage of traditional communication infrastructure during disasters, ensuring real-time and stable communication support. The network topology is dynamically adjusted to adapt to complex environments and changing needs, improving communication reliability. Through feature extraction and semantic analysis, image data is segmented and prioritized, enabling critical data to be transmitted to decision-makers or rescue teams at a faster speed, optimizing resource allocation and utilization. With a dual-loop feedback mechanism, the system can adjust data transmission strategies in real time according to changes in the communication environment, ensuring transmission efficiency and flexibility, and enhancing the system's ability to cope with different scenarios. This embodiment can acquire disaster information more quickly, analyze the disaster situation, and guide rescue operations, thereby improving the speed and accuracy of disaster response.

[0028] According to one aspect of this application, the steps of performing adaptive multimodal image acquisition include:

[0029] S11. Obtain disaster area information and plan UAV deployment. Read geographical and satellite coverage data of the disaster area, and combine this with meteorological data to construct a 3D disaster environment model. Based on the 3D disaster environment model, calculate the satellite communication blind zone distribution map (e.g., using ray tracing algorithms), dividing the blind zones into complete blind zones and edge blind zones. Based on the satellite communication blind zone distribution map, perform adaptive grid segmentation to generate a UAV coverage grid, with each grid containing priority and coverage difficulty coefficients. Based on the UAV coverage grid and currently available UAV resource parameters, perform optimization calculations to generate an initial UAV deployment plan, including the location and altitude information of each UAV. For the initial UAV deployment plan, verify the relay communication coverage rate. If the coverage rate is lower than a threshold, adjust the node positions and altitudes, and finally output an optimized UAV deployment plan.

[0030] S12. Configure and execute adaptive multimodal image acquisition. Based on disaster type data and optimized UAV deployment schemes, configure an image acquisition parameter set for each UAV, including resolution, shooting angle, exposure settings, and sampling frequency. According to the satellite communication blind zone distribution map, divide the image acquisition priority areas and assign different acquisition densities and priorities to different areas. Execute multimodal image acquisition, automatically switching between visible light, infrared, and multispectral modes based on real-time scene complexity scores (e.g., through an adaptive mode switching algorithm), generating a raw image dataset. Perform real-time quality assessment on the raw image dataset, calculate image quality indicators, and trigger a re-acquisition process for areas that do not meet quality standards. Attach metadata tags to each image, including spatial location, timestamp, acquisition mode, and quality indicators, generating a labeled image dataset.

[0031] This embodiment takes into account the impact of terrain undulations and buildings on satellite signals, and dynamically calculates blind zone changes in conjunction with meteorological factors (such as rainfall and smoke); it automatically switches imaging modes according to the real-time scene complexity, and introduces a quality assessment feedback mechanism to achieve closed-loop control of acquisition-assessment-reacquisition, thereby improving acquisition efficiency and image quality.

[0032] According to one aspect of this application, the steps for constructing and maintaining a relay network topology include:

[0033] S21. Establish a low-altitude UAV relay network. Based on a labeled image dataset, initialize each UAV node and establish an initial relay network topology, including the link relationships between nodes and communication quality indicators. Perform network connectivity testing to identify potential weak connections in the network and generate a network vulnerability map. Based on the network vulnerability map, apply a topology self-optimization algorithm to fine-tune the UAV positions and antenna pointing to generate an enhanced relay network topology. Configure the communication parameters of each node, including frequency, power, and coding scheme, and dynamically adjust them according to node position and link quality to generate relay node communication configurations. Deploy an improved distributed routing protocol and build a multi-path routing table to ensure the reliability and robustness of data transmission.

[0034] S22. Maintain the relay network topology and communication links. Continuously monitor real-time communication quality metrics for each relay node, including signal-to-noise ratio, bit error rate, latency, and throughput. Based on real-time communication quality metrics, construct a communication quality heatmap to dynamically display the communication status of different areas of the network. Analyze historical communication quality data (e.g., using predictive link degradation detection algorithms) to identify potential link degradation trends and generate link health warnings. Based on link health warnings and the communication quality heatmap, trigger a dynamic topology adjustment mechanism and update adjustment strategy instructions. Execute the adjustment strategy instructions, coordinating relevant UAVs to fine-tune their positions and parameters to maintain the stability and efficiency of the enhanced relay network topology.

[0035] This embodiment introduces the concept of a network vulnerability map, which improves the overall network robustness by proactively identifying and strengthening weak links in the network.

[0036] According to one aspect of this application, the steps for generating a link health warning include:

[0037] Multidimensional communication quality data preprocessing. Real-time communication quality indicators and stored historical communication quality data are read, and the data is organized by UAV node ID and time series, then timestamped to generate an aligned time-series dataset. Outlier detection is performed on the aligned time-series dataset, identifying and marking outlier data points. Outliers are replaced using local smoothing interpolation to obtain a cleaned time-series dataset. Multi-scale time window partitioning is performed on the cleaned time-series dataset, generating short-term (30 seconds), medium-term (5 minutes), and long-term (30 minutes) multi-scale time window data.

[0038] Link stability feature extraction. Statistical features, including mean, variance, kurtosis, skewness, and extreme values, are calculated for each window of the multi-scale time-window data to generate a basic statistical feature set. Wavelet transform (e.g., using a joint time-domain and frequency-domain analysis method) is applied to the cleaned time-series dataset to extract frequency-domain features, resulting in a frequency-domain feature vector. Trend indicators in the cleaned time-series dataset are calculated, including first-order differencing, second-order differencing, and linear regression slope, generating a trend feature vector. The correlation of communication quality between adjacent UAV nodes is analyzed, and a node correlation matrix is ​​constructed as a network-level feature. The basic statistical feature set, frequency-domain feature vector, trend feature vector, and node correlation matrix are integrated to form a complete link stability feature set.

[0039] Multimodal degradation pattern recognition. Typical degradation patterns, including signal attenuation, interference spikes, enhanced multipath effects, and antenna pointing offset, are loaded from a predefined link degradation pattern library. The link stability feature set is matched against the patterns in the link degradation pattern library, and a similarity score is calculated for each pattern (e.g., using a dynamic time warping algorithm), generating degradation pattern matching results. Based on the degradation pattern matching results, a fuzzy logic inference engine is applied for pattern fusion to handle complex situations where multiple degradation patterns coexist, outputting an evaluation of the current degradation pattern.

[0040] Degradation trend prediction based on residual networks. The process involves reading a cleaned time-series dataset and evaluating the current degradation mode to construct an input feature matrix, which is then fed into a pre-trained lightweight residual network model. The residual network model captures subtle trends in communication quality through a residual learning framework, predicting communication quality indicators (QIs) at future time points (1 minute, 3 minutes, and 5 minutes), generating QI prediction values. These QI prediction values ​​are then compared with a predefined safety threshold parameter table to calculate the degradation risk coefficient at each time point, generating a degradation risk time-series graph.

[0041] Generate adaptive threshold-based early warnings. For different UAV nodes and link types, read the basic early warning threshold from the system configuration parameters and, combined with historical early warning accuracy, adaptively adjust the threshold to generate customized early warning thresholds. Compare the degradation risk time series graph with the customized early warning thresholds, and determine the early warning level (minor, moderate, severe) based on the degree and duration of exceeding the threshold, generating a preliminary link early warning candidate set. Analyze the early warning status of adjacent links in the candidate set (e.g., applying spatial correlation filtering algorithms), eliminate isolated early warnings, enhance related early warnings, and output optimized link health early warnings. Add metadata to the link health early warnings, including the predicted degradation time, possible causes, suggested mitigation measures, and confidence scores, forming a complete early warning information package.

[0042] Early warning accuracy assessment and model update. Continuously record issued early warning packets and observed changes in communication quality, calculate early warning accuracy, recall, and F1 score, and generate an early warning effectiveness assessment. Based on the early warning effectiveness assessment, update the degradation prediction model parameters (e.g., by applying incremental learning algorithms) to improve prediction accuracy. Add newly identified degradation pattern features to the link degradation pattern library, expanding the library's coverage and enabling its self-evolution.

[0043] Traditional methods primarily rely on single indicators (such as signal-to-noise ratio degradation) to determine link status. This embodiment, however, introduces a comprehensive analysis of multi-dimensional features (time domain, frequency domain, network topology), enabling the identification of complex degradation patterns. It can also handle degradation processes at different rates, solving the problem that traditional fixed-window methods cannot adapt to varying degradation speeds. The residual learning framework focuses on learning the magnitude of changes in communication quality indicators rather than their absolute values, improving sensitivity to subtle trend changes. Traditional prediction methods often fail to react adequately to situations where absolute value changes are small but the trend is significant. The residual learning framework amplifies subtle changes, allowing the system to detect potential degradation 3-5 minutes in advance. Traditional fixed-threshold methods struggle to adapt to different link characteristics and environmental conditions. This embodiment introduces adaptive threshold adjustment based on historical accuracy and spatial correlation filtering, reducing false alarm rates (by approximately 40%) and false negative rates (by approximately 25%). It also considers the mutual influence between links in the network topology, distinguishing between isolated noise and systemic degradation trends.

[0044] According to one aspect of this application, the step of forming a transmission priority sequence includes:

[0045] S31. Perform image feature extraction and semantic analysis. Read the labeled image dataset, construct a multi-resolution image pyramid, and generate multi-scale image representations. Apply a lightweight disaster feature extractor to the multi-scale image representations to calculate basic feature vectors, including edge, texture, color, and shape features. Input the basic feature vectors into a disaster scene classification model to generate scene type probability distributions and disaster element identification results. Based on the disaster element identification results, perform semantic segmentation to generate a semantic segmentation map, marking key areas such as personnel, building damage, fire, and water areas. Calculate the correlation between regions in the semantic segmentation map, construct a semantic correlation graph, and describe the spatial and semantic relationships between different regions.

[0046] S32. Implement adaptive image segmentation and dynamic priority calculation. Based on the semantic segmentation map, apply an adaptive quadtree segmentation algorithm to generate non-uniform image segments, ensuring semantic integrity. Calculate a basic priority score for each non-uniform image segment, comprehensively considering disaster type, element importance, and area proportion to generate a basic segment priority. Based on the semantic association graph, assign association coefficients to interconnected segment groups, adjust the basic segment priorities, and generate association adjustment priorities. Integrate segment size data, compression difficulty coefficient, and association adjustment priorities, and apply a multi-factor weighted model to calculate transmission resource demand indicators. Based on the transmission resource demand indicators and association adjustment priorities, generate an initial transmission priority sequence (e.g., by applying a scheduling algorithm).

[0047] This embodiment uses an adaptive quadtree segmentation algorithm to dynamically adjust the segmentation boundaries based on semantic content, ensuring that key information is not fragmented, unlike traditional fixed grid segmentation methods. It introduces the concept of a semantic association graph to ensure that related content is transmitted in a proper order, avoiding the problem of isolated information being difficult to understand. The priority calculation model comprehensively considers three dimensions: information value, resource consumption, and semantic integrity, forming a multi-objective optimization framework, rather than a simple single-factor ranking.

[0048] According to one aspect of this application, the steps of generating non-uniform image patches include:

[0049] Perform semantic analysis on the labeled image dataset to generate a semantic segmentation map;

[0050] Based on the semantic segmentation map, semantic boundaries and importance features are extracted to generate a segmentation guidance map;

[0051] Based on the segmentation guidance map, an adaptive quadtree segmentation algorithm is applied to generate the initial fragmentation structure;

[0052] Perform semantic boundary-driven boundary adjustment on the initial fragmentation structure to generate a boundary-optimized fragmentation structure;

[0053] Semantic integrity protection processing is performed on the boundary-optimized segmentation structure to generate non-uniform image segments adapted to the characteristics of disaster scenarios.

[0054] According to one aspect of this application, the steps for performing semantic integrity protection processing on a boundary-optimized fragmentation structure include:

[0055] Perform fragment merging on the same semantic object that is split in the boundary-optimized fragmentation structure, and split the fragments that exceed the preset threshold after merging along the path with the minimum importance gradient to generate a semantically optimized fragmentation structure.

[0056] An adaptive scale balancing algorithm is applied to the semantically optimized fragmentation structure and isolated small fragments are processed. The merging strategy is determined based on semantic similarity and importance to generate a balanced fragmentation structure.

[0057] For high-value information regions in the balanced processing fragmentation structure, the fragment size is adjusted according to the characteristics of the transmission window, so that key information can be transmitted within a short transmission window, generating non-uniform image fragments.

[0058] Specifically, the semantic segmentation map is read, region boundary information is extracted, and the boundary clarity between semantic regions is enhanced (e.g., by applying the Canny edge detection algorithm), generating an enhanced boundary map. Combining the disaster element identification results, importance weights are assigned to each semantic region in the semantic segmentation map, transforming it into an importance weight map. The texture complexity and content richness of each region in the semantic segmentation map are calculated to generate a complexity map. The enhanced boundary map, importance weight map, and complexity map are fused, and a multi-layer weighted algorithm is applied to generate a comprehensive segmentation guidance map.

[0059] An initial quadtree segmentation is applied to the original image data. Based on the values ​​in the segmentation guidance map, it is determined whether further subdivision of the region is needed, generating a preliminary quadtree segmentation structure. The quadtree segmentation structure is evaluated to identify the following problems: Semantic fragmentation: Important semantic objects (such as people) are assigned to multiple different segments; Boundary misalignment: Segment boundaries do not match semantic boundaries; Scale imbalance: Regions of similar importance have large differences in segment scale due to differences in local features; Island effect: Small, highly important regions form isolated small segments. These problems are analyzed to generate a problem region labeling map, providing a basis for subsequent improvements.

[0060] Read the quadtree sharding structure and problem region labeling map to identify misaligned boundary areas. Apply a semantic boundary adjustment algorithm to move shard boundaries towards the nearest semantic boundary, resolving the boundary misalignment problem and generating a boundary-optimized sharding structure. Set a maximum boundary adjustment range threshold to ensure that the adjusted shards still conform to the basic constraints of the quadtree and avoid excessive distortion. Mark the small semantically incomplete regions that still exist in the boundary-optimized sharding structure to generate residual problem labels.

[0061] The algorithm reads the boundary optimization fragmentation structure and residual problem markers to identify semantically fragmented regions. For cases where the same semantic object is divided into multiple fragments, semantically driven fragment merging is performed to generate a semantically merged fragmentation structure. For fragments that are too large after merging, importance-aware secondary segmentation is performed, segmenting along the path with the least importance gradient to ensure that the segmentation line does not cross important semantic regions as much as possible, resulting in an optimized segmentation structure. The semantic integrity score of each fragment is calculated and recorded in the fragment semantic integrity table.

[0062] The optimized segmentation structure is read, and the fragment size distribution is analyzed to identify scale-uneven regions and isolated regions. An adaptive scale equalization algorithm is applied to dynamically adjust the fragmentation threshold based on regional importance, resolving the scale imbalance problem and generating a scale-balanced fragmentation structure. For the isolated effect, an intelligent merging strategy is applied: if an isolated small fragment has high semantic similarity to its neighboring fragments, it is merged into the neighboring fragment; if the semantic similarity is low but the importance is extremely high, the independent fragment is retained; if the semantic similarity is low and the importance is moderate, the decision to merge is made based on transmission resource conditions. Executing the above strategy generates an isolated fragmentation structure.

[0063] This paper analyzes the adaptability of fragmentation in the isolated image processing fragmentation structure under different transmission conditions, considering fragment size, content complexity, and compression efficiency. For the special requirements of disaster scenarios (such as priority transmission of emergency information and limited transmission windows), a transmission window-aware fragmentation optimization algorithm is applied to make final adjustments to the fragmentation structure. For fragments containing high-value information such as personnel and hazards, the paper ensures that their size is suitable for transmission within a short time window, and further subdivides them if necessary. The final non-uniform image fragmentation structure is generated, and each fragment is associated with its semantic label, importance score, and estimated transmission resource requirements.

[0064] A comprehensive evaluation is performed on non-uniform image slices, calculating the following metrics: Semantic integrity score: the completeness of semantic objects within the slice; Boundary alignment: the degree of conformity between the slice boundary and the semantic boundary; Scale uniformity: the uniformity of slice size distribution; Transmission adaptability: the suitability of the slice under different transmission conditions. A complete slice quality evaluation report is generated, recording each metric. Detailed metadata is generated for each slice, including location, size, contained semantic objects, importance score, and suggested transmission priority, forming a slice metadata table. The final non-uniform image slices and their associated slice metadata tables are output for subsequent priority calculation and transmission scheduling.

[0065] Traditional adaptive quadtrees primarily determine segmentation based on color and texture differences, while this embodiment introduces semantic boundaries as the dominant factor. In disaster scenarios, semantic information (personnel location, danger zone) is more important than texture information, ensuring that segmentation boundaries align with semantic boundaries and preventing critical information from being segmented into different segments. Conventional quadtrees cannot understand the actual importance of content, while this embodiment uses the importance of disaster elements as the core decision factor. In particular, for high-priority elements such as personnel and hazard sources, their integrity and independence are ensured even if their area is small, while large areas of low priority (such as open backgrounds) are allowed to be merged to a greater extent. This asymmetric processing is particularly suitable for the uneven value distribution of information transmission at disaster sites. Traditional methods usually only consider content characteristics to determine segmentation, ignoring transmission condition constraints. This embodiment incorporates transmission window characteristics into the segmentation decision process, ensuring that the segment size of high-value information is suitable for short-term transmission windows, while low-value information can be divided into larger segments to improve overall efficiency, achieving synergistic optimization of segmentation structure and transmission conditions. In response to the extremely uneven distribution of information value in disaster scenarios, this embodiment introduces a coarse-grained and fine-grained two-layer fragmentation strategy. Conventional fragmentation is used for general areas, while finer-grained fragmentation is used for high-value areas (especially areas that may change rapidly, such as the front line of a fire or areas where people gather). The correlation between fine-grained fragments is preserved, which solves the contradiction that traditional single-scale fragmentation cannot simultaneously satisfy semantic integrity and transmission flexibility.

[0066] According to one aspect of this application, the step of calculating the transmission priority of each fragment includes:

[0067] Calculate the disaster type, element importance, and area proportion of each non-uniform image patch to generate the basic priority of the patch;

[0068] Read the semantic segmentation map and basic segmentation priority, identify segment pairs with spatial adjacency and semantic continuity, measure the degree of semantic association by calculating the cosine similarity of semantic feature vectors, and generate an association strength matrix;

[0069] Based on the correlation strength matrix, fragments with correlation strength exceeding a preset threshold are organized into priority groups, and an overall transmission sequence number is assigned to each group to generate a priority group table.

[0070] Within each priority group, the relative priority within the group is calculated based on the ratio of the basic priority of the fragments. The overall transmission sequence number of the group is then combined with the relative priority within the group to form the associated adjustment priority, i.e., the transmission priority, to ensure the coherent transmission of semantically associated fragments.

[0071] According to one aspect of this application, performing semantic integrity protection processing on the boundary-optimized fragmentation structure further includes:

[0072] Spatial overlay analysis is performed on the boundary-optimized piecewise structure and the semantic segmentation map to calculate the proportion of each semantic object being segmented into different pieces. When the segmentation proportion exceeds a preset threshold, it is marked as a semantic fragmentation region. Semantic fragmentation region labels are generated and semantic fragmentation loss is calculated.

[0073] For each semantic object in the semantically fragmented region label, a fragmentation merging cost function is constructed by combining the preset fragmentation boundary importance score. When the fragmentation merging cost is lower than the semantic fragmentation loss, fragmentation merging is performed to generate a semantically merged fragmentation structure.

[0074] Analyze the size distribution of each piece in the semantic merging fragmentation structure, calculate the coefficient of variation of the fragment size and the distribution of the number of fragments in each size interval, and generate a scale distribution map;

[0075] By combining a pre-defined importance map and scale distribution map, an adaptive scale adjustment function is constructed. More granular segmentation thresholds are applied to high-importance areas, while larger-size segments are allowed for low-importance areas, generating a scale-balanced segmentation structure that achieves a balance between the value of disaster information and the demand for transmission resources.

[0076] Specifically, the difference between the segmentation ratio of a semantic object in different segments and a preset threshold is quantified. The larger the difference, the more severe the semantic fragmentation. This quantified difference is multiplied by a preset weight coefficient to obtain the semantic fragmentation loss. For example, if the maximum segmentation ratio of a semantic object in different segments is p, the preset threshold is t, and the weight coefficient is w, then the semantic fragmentation loss L1 = w × (pt) (when p > t).

[0077] According to one aspect of this application, generating non-uniform image patches adapted to the characteristics of disaster scenarios further includes:

[0078] Analyze the content types and disaster risk assessment results in the semantic segmentation map, divide the image region into high dynamic region and low dynamic region, and generate a region type labeling map;

[0079] Fine-grained segmentation parameters are applied to high-dynamic regions in the region type labeling map, and coarse-grained segmentation parameters are applied to low-dynamic regions, forming a two-layer segmented initial structure containing fine-grained and coarse-grained layers.

[0080] Establish the spatial correspondence between fine-grained and coarse-grained segments in the initial two-layer segmentation structure, construct an inter-layer mapping table, and record the coarse-grained segment number and its relative position to which each fine-grained segment belongs.

[0081] The consistency of the two-layer sharding boundary is optimized based on the inter-layer mapping table to ensure that the fine-grained layer sharding boundary does not cross the coarse-grained layer sharding boundary, thus generating an optimized two-layer sharding structure.

[0082] During transmission, based on the transmission window status and regional change monitoring results, the system dynamically selects whether to use fine-grained or coarse-grained sharding layers to achieve precise transmission of high-value, rapidly changing areas and efficient transmission of general areas.

[0083] According to one aspect of this application, the step of performing adaptive transmission of image fragments includes:

[0084] S41. Monitor and predict the communication environment and resource status. Each UAV node collects real-time communication parameters, including signal strength, bit error rate, available bandwidth, latency, and jitter. Combining real-time communication parameters and historical communication data, a time-series prediction model is applied to generate short-term communication quality predictions, forecasting communication status changes over the next 30 seconds to 5 minutes. Based on the short-term communication quality predictions and UAV node locations, a transmission window model is constructed to determine the available time window and bandwidth capacity for each transmission path. The computing and storage resource status of each UAV node is monitored, and combined with power monitoring data, a resource availability report is generated. The transmission window model and resource availability report are integrated to construct a comprehensive resource status map, serving as the basis for scheduling decisions.

[0085] S42. Implement dual-loop feedback transmission scheduling and execution. Read the initial transmission priority sequence and comprehensive resource status diagram, apply the time-value-size three-dimensional optimization model, and generate a fine-grained scheduling plan. For high-priority fragments in the fine-grained scheduling plan, configure adaptive encoding parameters, select differentiated encoding schemes for different content types, and generate encoding configuration instructions. Execute a fast loop feedback mechanism, fine-tuning transmission parameters and updating transmission control instructions every 50 milliseconds based on real-time communication status. Execute a slow loop feedback mechanism in parallel, updating the fine-grained scheduling plan and optimizing medium- and long-term transmission strategies every 500 milliseconds based on transmission confirmation information. Execute image fragment transmission according to transmission control instructions, record transmission process data, and generate a transmission execution log. Continuously analyze the transmission execution log, evaluate transmission efficiency and key information coverage, and generate a transmission effect evaluation report. Based on the transmission effect evaluation report, update the scheduling model parameters and simultaneously send feedback information to steps S22 and S32 to achieve system-level optimization.

[0086] This embodiment separates real-time fine-tuning from strategy adjustment through a dual-loop feedback mechanism. The fast loop ensures immediate response to communication fluctuations, while the slow loop guarantees stable optimization of the overall transmission strategy. A three-dimensional optimization model dynamically balances three key factors—time, value, and size—and employs differentiated strategies at different transmission window stages to maximize transmission value. The differentiated coding scheme automatically selects appropriate coding parameters based on content type (e.g., people, buildings, background), unlike traditional uniform coding strategies. An active learning mechanism during transmission extracts patterns from historical transmission results, continuously optimizing the scheduling model and enabling the system to adapt.

[0087] like Figure 2 As shown, according to one aspect of this application, the steps of dynamically adjusting the transmission strategy through a dual-loop feedback mechanism include:

[0088] Read the transmission priority sequence and communication environment status, apply the time-value-size three-dimensional optimization model, and generate a fine-grained scheduling plan, i.e., a transmission strategy;

[0089] Real-time communication status is collected at the first preset period, and transmission control commands are generated in combination with a fine scheduling plan.

[0090] The transmission confirmation information is collected in the second preset cycle to evaluate the execution effect of the transmission control command and update the fine scheduling plan.

[0091] The first preset cycle is shorter than the second preset cycle, forming a dual-loop feedback structure for rapid parameter adjustment and strategy optimization.

[0092] like Figure 3 As shown, according to one aspect of this application, the steps for generating a fine-grained scheduling plan by applying a time-value-size three-dimensional optimization model include:

[0093] Read the transmission priority sequence and communication environment status, construct a three-dimensional parameter space, and generate a fragmentation basic parameter table and transmission window characteristic description; based on this, divide the transmission window into three stages: early, middle and late, and set differentiated weight configurations for each stage;

[0094] Based on the basic parameter table for segmentation, the value of segmented content, the correlation adjustment value, and the timeliness coefficient are calculated to generate a comprehensive value score.

[0095] Based on the characteristics of the transmission window, a transmission risk assessment is performed, and a time risk matrix is ​​generated.

[0096] Based on the analysis of the fragmentation basic parameter table, transmission resource efficiency is determined, and a size adjustment strategy is generated.

[0097] The comprehensive value score, time risk matrix, and size adjustment strategy are input into the optimization objective function to obtain a fine-grained scheduling plan. In the early stage of the transmission window, high-value, large-volume fragments are prioritized for transmission, while in the late stage of the transmission window, high-value, small-volume fragments are prioritized for transmission.

[0098] Specifically, the initial transmission priority sequence is read, and the basic priority value, slice size, and estimated transmission time of each image slice are extracted to construct a slice basic parameter table. The estimated start time, expected duration, and bandwidth fluctuation pattern of the current transmission window are extracted from the comprehensive resource state diagram to generate a transmission window characteristic description. Based on the transmission window characteristic description, the transmission window is divided into three stages: early, middle, and late. Different optimization weight parameters are set for each stage to form a stage weight configuration. The system's preset optimization parameter template is read, and combined with the current task type and transmission conditions, the model configuration parameters are initialized, including the value decay coefficient, size penalty factor, and time window risk coefficient.

[0099] For each segment in the basic parameter table, its semantic tags and content type information are read. Combined with the disaster emergency decision-making table, a basic value score related to the content is calculated. Considering the semantic relationships between segments, a semantic relationship graph is read. For strongly related segment groups, the synergistic value gain within the group is calculated, and the relationship adjustment value of each segment is updated. The timeliness characteristics of the segment content are analyzed. For highly time-sensitive content (such as personnel location, the forefront of fire spread, etc.), a time decay function is applied to calculate the timeliness adjustment coefficient. Combining the basic value score, relationship adjustment value, and timeliness adjustment coefficient, a comprehensive value score for each segment is calculated.

[0100] Based on the characteristics of the transmission window, a Monte Carlo simulation method is used to perform probabilistic analysis on the reliability and duration of the transmission window, generating a window reliability distribution. For each fragment, based on its size and current transmission conditions, the probability distribution of its transmission completion time is estimated, generating a transmission time distribution. A transmission risk assessment function is designed, combining the window reliability distribution and the transmission time distribution to calculate the transmission risk coefficient of each fragment at different scheduling time points, generating a time risk matrix. For the later stages of the transmission window, risk assessment weights are increased to ensure that critical small-volume fragments are transmitted first, updating the corresponding values ​​in the time risk matrix.

[0101] Analyzing current available bandwidth and expected fluctuation patterns, and combining this with the fragmentation size from the fragmentation basic parameter table, a transmission efficiency coefficient is calculated to generate a transmission efficiency assessment. Based on a dynamic programming algorithm, the resource utilization efficiency of different fragmentation combinations under limited bandwidth conditions is calculated, identifying the optimal fragmentation combination and forming an optimal resource combination. Considering transmission interruption recovery costs, for large fragments, their transmission reliability and recovery overhead under different network conditions are evaluated, and a size penalty coefficient is calculated. In the early stages of the transmission window, the size penalty weight is reduced to improve bandwidth utilization; in the later stages of the window, the preference for small, high-value fragments is increased, generating a phased size adjustment strategy.

[0102] A three-dimensional optimization objective function is constructed, taking the comprehensive value score, time risk matrix, and size penalty coefficient as inputs to form the optimization objective function. Constraints are set, including total transmission time not exceeding the window duration, priority-based dependency enforcement, and resource utilization lower limits, generating a constraint set. An optimal scheduling scheme is searched in the three-dimensional parameter space (e.g., using particle swarm optimization), generating a preliminary optimized scheduling scheme. The critical path and bottlenecks in the optimized scheduling scheme are analyzed, and a local search algorithm is applied for fine-tuning to improve the scheme's robustness, resulting in a preliminary draft of a refined scheduling plan.

[0103] Based on the three phases (early, mid, and late) of the transmission window, different strategy weights are assigned to each phase in the initial draft of the fine-grained scheduling plan: Early phase strategy: favors high-value, large-volume fragments to maximize bandwidth utilization and information transmission; Mid-phase strategy: balances value, risk, and size, executing according to the optimization model results; Late-phase strategy: favors high-value, small-volume fragments to maximize the success rate of critical information transmission. A smooth transition mechanism is designed between phases to avoid resource waste caused by sudden policy changes, generating phase transition functions. An emergency backup plan is added to quickly switch to an emergency mode prioritizing critical information in the event of an abnormally shortened transmission window, forming an emergency scheduling strategy. The main strategy and emergency strategy are integrated to generate a complete fine-grained scheduling plan, including the transmission order, time window, and scheduling priority of each fragment.

[0104] Traditional optimization models typically use fixed weights, while this embodiment divides the transmission window into three stages, dynamically adjusting the weights of value, time risk, and fragment size in each stage. Especially in the later stages of the window, the system significantly increases the weight of small, high-value fragments, ensuring critical information is transmitted before the window closes. This "wide at the beginning, narrow at the end" transmission strategy is particularly suitable for unstable transmission environments at the edge of satellite communication blind spots. This embodiment also overcomes the limitation of traditional methods that calculate fragment value independently by introducing semantic correlation analysis between fragments. For example, multiple fragments containing images of the same disaster area from different perspectives, or a group of fragments showing the start and end points of a rescue channel, have a combined value greater than the sum of their individual values. Coordinated value gain ensures the continuity of transmission for semantically related fragments. A probabilistic model is introduced to assess transmission risk, rather than simply using deterministic estimation. Monte Carlo simulation considers random factors such as communication link fluctuations, bandwidth changes, and transmission interruptions, providing a more accurate probabilistic assessment of the transmission risk of each fragment at different time points, improving the reliability of decision-making in unstable communication environments.

[0105] like Figure 4 As shown, according to one aspect of this application, the step of generating transmission control instructions includes:

[0106] Real-time communication status is collected, noise is eliminated by Kalman filtering to generate filtered communication parameters, and short-term parameter change trends are predicted based on these parameters to construct a real-time communication status model.

[0107] Based on the real-time communication state model, the matching strategy is queried from the transmission parameter adjustment rule base, the fuzzy control algorithm is applied to determine the adjustment intensity, and the parameter adjustment amount and smoothing adjustment amount are generated.

[0108] By combining the fine-grained scheduling plan, the smooth adjustment amount is applied to the current transmission parameters to generate updated parameter values ​​and anomaly response instructions, thus forming transmission control instructions.

[0109] According to one aspect of this application, the step of generating the smoothing adjustment amount includes:

[0110] Read parameter adjustment amounts and historical adjustment records, calculate the frequency and magnitude of parameter adjustment direction changes within a predetermined period, and generate adjustment trend indicators;

[0111] Based on the adjustment trend indicator, potential parameter oscillation risks are identified, and oscillation risk scores for each transmission parameter are calculated.

[0112] Based on the oscillation risk score, an adaptive damping function is dynamically constructed. Oscillation risk scores exceeding the threshold are treated with a strong damping coefficient, while oscillation risk scores below the threshold are treated with a weak damping coefficient, generating a parameterized damping coefficient table.

[0113] The parameter adjustment amount is weighted and calculated with the corresponding damping coefficient in the parameterized damping coefficient table. The maximum amplitude and frequency of parameter adjustment are limited within a preset time to generate a smooth adjustment amount.

[0114] According to one aspect of this application, the steps for updating a fine-grained scheduling plan include:

[0115] Collect transmission confirmation information and fast loop operation logs, construct a comprehensive performance evaluation, reassess the remaining time and reliability of the transmission window, generate window status updates, detect changes in the transmission window stage to generate stage switching signals, and analyze bottlenecks in the transmission process to generate performance bottleneck reports.

[0116] Based on window status updates, phase switching signals, and performance bottleneck reports, formulate a transmission strategy adjustment plan and generate strategy adjustment recommendations.

[0117] Read the current fine-grained scheduling plan, update the scheduling order and time allocation of each segment according to the strategy adjustment suggestions, and generate an updated scheduling plan;

[0118] By combining the updated scheduling plan and transmission control instructions, the resource allocation and encoding strategies for subsequent fragments are adjusted to generate a resource optimization scheme. This resource optimization scheme is then integrated with the updated scheduling plan into a refined updated scheduling plan. Specifically, when a transmission window is detected as potentially closing prematurely, the system automatically switches to an emergency transmission mode prioritizing critical information.

[0119] Specifically, real-time communication status parameters, including key indicators such as signal-to-noise ratio, bit error rate, current bandwidth, and link latency, are collected from the communication module at 50-millisecond intervals. Kalman filtering is applied to the collected real-time communication status parameters to eliminate instantaneous noise and outliers, resulting in smoothed filtered communication parameters. The short-term rate of change and trend of key parameters are calculated to predict parameter changes in the next 100-200 milliseconds, generating short-term trend predictions. Combining the filtered communication parameters and short-term trend predictions, a real-time communication status model is constructed as the basis for fast-loop decision-making.

[0120] Based on the real-time communication state model, matching adjustment strategies are queried from the transmission parameter adjustment rule base to generate a preliminary adjustment plan. The adjustment magnitude (e.g., by applying a fuzzy control algorithm) is determined based on the parameter change amplitude and rate, generating the parameter adjustment amount. To avoid parameter oscillations, a damping mechanism is introduced to limit the maximum amplitude and frequency of parameter adjustments in the short term, updating the smooth adjustment amount. Combining the current transmission fragmentation type and importance, differentiated adjustment strategies are read from the fragmentation adaptive configuration to further optimize the smooth adjustment amount.

[0121] The smooth adjustment is applied to the current transmission parameters, including coding rate, packet size, retransmission timeout threshold, and forward error correction level, to calculate updated parameter values. Specific response strategies are applied for different types of communication anomalies (such as sudden interference or persistent attenuation), generating anomaly response instructions. The updated parameter values ​​and anomaly response instructions are organized into a standard format transmission control command and sent to the transmission execution module. Detailed information for each parameter adjustment and command transmission is recorded, including the reason for the adjustment, the adjustment amount, and the timestamp, forming a fast-loop operation log.

[0122] Collect transmission confirmation information every 500 milliseconds, including completed fragmentation, transmission success rate, actual transmission rate, and resource utilization efficiency. Read the fast loop operation log from the past 10 seconds, analyze the effect and trend of fast loop adjustments, and generate a fast loop effect evaluation. Retrieve the status information of fragments to be transmitted from the queue, and combine it with the updated data of the comprehensive resource status graph to generate a transmission task status report. Integrate the transmission confirmation information, fast loop effect evaluation, and transmission task status report to construct a comprehensive performance evaluation.

[0123] Based on the transmission rate and success rate data from the comprehensive performance evaluation, the transmission window model is updated, the remaining window time and reliability are re-estimated, and a window state update is generated. According to the window state update, it is checked whether the current transmission window stage (early, middle, or late) needs to be switched early, and a stage switching signal is generated. Abnormal patterns and bottlenecks in the transmission process are analyzed, identifying key areas requiring optimization, and a performance bottleneck report is generated. Combining the window state update, stage switching signal, and performance bottleneck report, a transmission strategy adjustment plan is formulated, and strategy adjustment recommendations are generated.

[0124] The system reads the existing fine-grained scheduling plan and, based on policy adjustment suggestions, updates the scheduling order and time allocation of each fragment, generating an updated scheduling plan. For cases of significant changes in the transmission window state (e.g., a window shortening of more than 50%), an emergency scheduling strategy is triggered to reorder the remaining fragments, prioritizing the transmission of high-value content. Based on actual transmission progress and resource consumption, the resource allocation and encoding strategies for subsequent fragments are adjusted, generating a resource optimization scheme. The updated scheduling plan and resource optimization scheme are integrated into a complete policy update package, outputting the updated fine-grained scheduling plan.

[0125] Design a data sharing interface between the fast and slow loops to ensure that the fast loop can access the strategy decisions of the slow loop, and the slow loop can analyze the operation records of the fast loop. Implement a priority-based interrupt mechanism, allowing the slow loop to immediately notify the fast loop to adjust its strategy when it detects significant changes in the transmission environment, avoiding policy delays. Establish a clock synchronization mechanism between the loops to ensure that the time bases of the two loops are consistent, facilitating subsequent analysis and evaluation. Maintain a unified state database to record the operation history and effects of the two loops, supporting system self-learning and optimization.

[0126] Traditional feedback control systems typically employ a single cycle, making it difficult to simultaneously meet the demands of real-time response and strategy optimization. This embodiment separates the cycle into a 50-millisecond fast cycle and a 500-millisecond slow cycle, analogous to the reflex arc and central nervous system in biological systems. The fast cycle focuses on immediate fine-tuning of communication parameters, ensuring rapid response to sudden disturbances; the slow cycle handles strategy-level optimization, adjusting transmission plans and resource allocation, thus improving the system's adaptability to communication changes across different time scales. Unlike traditional purely reactive adjustments, the fast cycle introduces a short-term prediction mechanism, predicting the state 200 milliseconds into the future based on the changing trends of communication parameters, achieving "proactive response" rather than "passive reaction." Simultaneously, a damping control mechanism is introduced to prevent excessive adjustment and oscillation of parameters in environments with severe fluctuations, improving the stability and robustness of the control system. Breaking through the limitations of traditional fixed strategies, the slow cycle can not only adjust transmission parameters but also dynamically switch transmission stages and strategies based on the transmission window state. For example, when it detects that the transmission window may close prematurely, the system automatically switches from a "maximum bandwidth" strategy to a "critical information priority" strategy, ensuring the transmission of the highest-value information within a limited time.

[0127] Case 1: In a disaster relief scenario following an earthquake in a mountainous area, a relay network consisting of four drones was used to efficiently transmit images of edge areas via satellite communication. A specific method for acquiring and transmitting images of disaster sites in satellite communication blind spots using low-altitude drone relay is as follows:

[0128] Step 1: Implementation of disaster environment perception and image acquisition.

[0129] The system first acquires disaster area information, including topographic data (DEM, Digital Elevation Model), satellite coverage parameters, and meteorological conditions. Based on this data, a satellite communication blind zone distribution map is constructed. The satellite communication blind zone calculation method is as follows: Blind zone probability P(x, y) = Ps·Pm·Pt; where Ps is the probability of terrain obstruction, determined by calculating the number of intersections between the terrain and the satellite using ray tracing; Pm is the probability of meteorological influence, calculated from the rainfall rate R (mm / h) as Pm = 1 - exp(-αR), where α is the meteorological attenuation coefficient; Pt is a time-related factor, calculated based on the visible time percentage according to satellite orbit parameters; and (x, y) represents the coordinates within the disaster area.

[0130] The system adaptively configures image acquisition parameters based on the blind zone distribution map, including the resolution selection function R(d) = Rmax·(1-β·d); where R is the acquisition resolution; Rmax is the maximum resolution parameter, typically 4096×3072 pixels; d is the normalized distance to the edge of the blind zone; and β is the resolution attenuation coefficient, typically 0.6.

[0131] To address scene complexity, the system employs an adaptive mode switching mechanism. The scene complexity score SC = w1·E + w2·V + w3·C; where E is the image entropy; V is the region variance; C is the edge density; and w1, w2, and w3 are weighting coefficients, set to 0.5, 0.3, and 0.2 respectively in implementation. When SC exceeds the threshold of 7.5, the system automatically switches from visible light mode to multispectral mode to improve information acquisition capabilities in special scenarios.

[0132] Step 2: Construction of UAV relay network and implementation of predictive link degradation detection.

[0133] The system establishes a low-altitude relay network consisting of four UAVs, initially deployed at the edge of communication blind spots to construct the network topology. During implementation, the system employs a predictive link degradation detection algorithm for link status monitoring.

[0134] Link stability feature extraction method: Link feature vector F = [Fm, Fv, Fs, Fp, Ft]; where Fm is the mean feature group, containing the average values ​​of parameters such as signal-to-noise ratio (SNR) and bit error rate (BER); Fv is the variance feature group, representing the degree of parameter fluctuation; Fs is the spectral feature, extracting frequency domain features through wavelet transform; Fp is the phase feature, capturing periodic change patterns; and Ft is the trend feature, containing first-order difference D1 and second-order difference D2.

[0135] Multimodal degradation pattern recognition employs the Dynamic Time Warping (DTW) algorithm to calculate the similarity between the feature sequence and typical degradation patterns: S(F, Mi) = DTW(F, Mi) / (1+λ·L); where S is the similarity score; F is the current link feature vector; Mi is the feature representation of the i-th predefined degradation pattern; DTW is the dynamic time warping distance; L is the sequence length; and λ is the length penalty coefficient, with a value of 0.05.

[0136] The core of degradation trend prediction based on residual networks lies in the residual learning framework. The prediction function is: Q*(t+Δt) =Q(t) + ΔQ(t, Δt); where Q* is the communication quality index at the predicted future time point; Q(t) is the communication quality index at the current time point; ΔQ is the change predicted by the residual network; t is the current time; Δt is the prediction time interval, and short, medium and long-term prediction values ​​are calculated by taking 1 minute, 3 minutes and 5 minutes respectively.

[0137] In implementation, the system uses an adaptive threshold for link early warning: Th(i,j) = Thbase(i)·(1-γ·Acc(j)); where Th is the customized early warning threshold; Thbase is the basic early warning threshold; i is the link type index; j is the UAV node number; Acc is the historical early warning accuracy; and γ is the accuracy adjustment coefficient, with a value of 0.3. When the degradation risk exceeds the threshold, the system triggers topology adjustment to ensure communication stability.

[0138] Step 3: Implementation of semantic-aware image processing with two-layer segmentation.

[0139] The system performs semantic analysis on the collected disaster images to generate semantic segmentation maps and identify key elements such as people, building damage, and fires. Based on this, the system adopts a two-layer segmentation strategy.

[0140] Regional Dynamic Characteristics Assessment: The system classifies image regions into high-dynamic and low-dynamic regions. The dynamic score DA(r) = wH·H(r) + wC·C(r) + wT·T(r); where DA is the dynamic score of region r; H is the disaster risk index, such as the speed of fire spread; C is the content change rate, calculated by the difference between adjacent time frames; T is the information timeliness requirement; wH, wC, and wT are weighting coefficients, which are taken as 0.4, 0.4, and 0.2 in implementation. When the dynamic score DA of region r exceeds the threshold of 8.0, it is marked as a high-dynamic region.

[0141] Dual-layer segmentation parameter settings: For high dynamic range regions, the system sets fine-grained parameters Pf = {dmin_f, dmax_f, th_f}; for low dynamic range regions, it sets coarse-grained parameters Pc = {dmin_c, dmax_c, th_c}; where dmin is the minimum segment size; dmax is the maximum segment size; and th is the splitting threshold. In implementation, Pf = {32, 128, 0.15}, Pc = {128, 512, 0.25}, in pixels.

[0142] Inter-layer relationship construction: The system establishes a mapping relationship between fine-grained shards and coarse-grained shards, with the mapping function M(i,j) = {c(i), pos(j,i)}; where i is the coarse-grained shard index; j is the fine-grained shard index; c(i) is the coarse-grained shard number containing fine-grained shard j; and pos(j,i) is the relative position encoding of fine-grained shard j within coarse-grained shard i.

[0143] Boundary consistency optimization: The system adjusts the fragmentation boundaries to ensure that fine-grained layer fragments do not cross coarse-grained layer boundaries. The optimization function is: B(e) = min(D(e, Ec)); where B is the boundary adjustment function; e is the fine-grained fragmentation boundary; D is the shortest distance function from a point to a line segment; and Ec is the set of coarse-grained fragmentation boundaries. Boundary adjustment is performed when the distance between the fine-grained fragmentation boundary and the coarse-grained fragmentation boundary exceeds a threshold of 10 pixels.

[0144] Step 4: Implementation of semantic association priority calculation.

[0145] The system calculates transmission priority based on non-uniform image segmentation, with the core being the consideration of semantic relationships between segments.

[0146] Piece base priority calculation: P_base(i) = wT·T(i) + wI·I(i) + wA·A(i); where P_base is the base priority of piece i; T is the disaster type weight of the piece; I is the element importance score; A is the area proportion factor; wT, wI, and wA are weight coefficients, which are 0.5, 0.3, and 0.2 respectively.

[0147] Semantic association strength calculation: C(i,j) = cos(V(i),V(j))·S(i,j); where C is the association strength between segments i and j; V is the semantic feature vector of the segment, obtained by dimensionality reduction of the semantic segmentation result; cos is the cosine similarity function; S is the spatial proximity index, which is 1 for adjacent segments, otherwise it is decayed by distance.

[0148] Priority group formation: The system organizes fragments with an association strength exceeding the threshold of 0.75 into priority groups, with the group transmission sequence number G(k) = max(P_base(i)), i∈gk; where G is the overall transmission sequence number of group k; gk is the kth fragment group; and max is the maximum value function.

[0149] Relative priority within a group: R(i, k) = P_base(i) / sum(P_base(j)), j∈gk; where R is the relative priority of fragment i within group k; sum is the summation function. Final association adjustment priority P(i) = G(k) + σ·R(i, k); where P is the final priority of fragment i; σ is the intra-group adjustment coefficient, taking a value of 0.1 to ensure coherent transmission of fragments within the same group.

[0150] Step 5: Implement dual-loop feedback transmission scheduling.

[0151] The system employs a dual-loop feedback mechanism for transmission scheduling, including a fast loop with a 50-millisecond cycle and a slow loop with a 500-millisecond cycle.

[0152] The time-value-size three-dimensional optimization model is implemented as follows: The system divides the transmission window into three stages: early, middle, and late, and applies different weight configurations: W(t) = {wV(t), wT(t), wS(t)}; where wV is the weight of the value dimension; wT is the weight of the time risk dimension; wS is the weight of the fragment size dimension; and t is the stage identifier of the transmission window. In the early stage (t=1), W(1) = {0.5, 0.2, 0.3}; in the middle stage (t=2), W(2) = {0.4, 0.4, 0.2}; and in the late stage (t=3), W(3) = {0.6, 0.3, 0.1}.

[0153] Comprehensive value assessment: V(i) = Vb(i) + α·Vr(i) + β·Ve(i); where V is the comprehensive value score of segment i; Vb is the basic value score; Vr is the correlation-adjusted value; Ve is the timeliness-adjusted value; α is the correlation value coefficient, with a value of 0.3; β is the timeliness coefficient, with a value of 0.4.

[0154] The transmission risk assessment employs the Monte Carlo simulation method: R(i,t) = 1 - Pr(Tc(i) < Tw(t)); where R is the transmission risk coefficient of fragment i at time t; Tc is the random variable of fragment transmission completion time; Tw is the random variable of the remaining time of the transmission window; and Pr is the probability calculation function. The system executes 1000 simulations to generate the risk probability distribution.

[0155] Fast Cyclic Damping Mechanism Implementation: The system calculates the oscillation risk of the parameter adjustment amount. The oscillation detection function is Osc(p) = count(sign(ΔP(tk, t-k+1)) ≠ sign(ΔP(t-k+1, t-k+2))) / (n-1); where Osc is the oscillation risk score of parameter p; ΔP is the parameter adjustment amount; sign is the sign function; count is the counting function; n is the observation window size, with a value of 3. Dynamic damping coefficient calculation is D(p) = Dmin + (Dmax - Dmin)·Osc(p)η; where D is the damping coefficient of parameter p; Dmin is the minimum damping coefficient, with a value of 0.2; Dmax is the maximum damping coefficient, with a value of 0.8; η is the nonlinear adjustment exponent, with a value of 2.0. Smoothing adjustment amount calculation is ΔP'(t) = ΔP(t)·(1-D(p)); where ΔP' is the smoothed parameter adjustment amount; ΔP is the original parameter adjustment amount.

[0156] In the practical application of mountain earthquake rescue scenarios, this embodiment shows significant advantages compared with traditional methods: improved transmission efficiency: the success rate of key information transmission is increased by 35%, and the average transmission delay is reduced by 42%; optimized resource utilization: the utilization rate of communication bandwidth is increased by 28%, and the energy efficiency of drones is increased by 25%; maximized information value: within a limited transmission window, the coverage rate of high-value information is increased by 56%; enhanced system robustness: in an environment with fluctuating communication quality, the link interruption recovery time is reduced by 68%. This embodiment realizes the efficient image acquisition and transmission in the satellite communication blind area of the disaster site through technologies such as a double-loop feedback mechanism, three-dimensional optimization of time-value-size, and a two-layer fragmentation strategy, providing reliable image information support for emergency rescue decision-making.

[0157] Case 2: Taking another mountain earthquake rescue scenario as the background, 4 drones are deployed to form a relay network to achieve the efficient acquisition and transmission of images at the edge of the satellite communication blind area. The method for acquiring and transmitting images of the satellite communication blind area at the disaster site by low-altitude drone relay is as follows:

[0158] S1. Obtain disaster area information and perform adaptive multimodal image acquisition.

[0159] First, obtain the basic information of the disaster area, including terrain data (DEM), satellite coverage data, and meteorological conditions, and then perform image acquisition.

[0160] S11. Terrain data processing and blind area calculation.

[0161] The system reads the mountain terrain DEM data (resolution 30m) and obtains the meteorological conditions (rainfall rate R = 8.5mm / h). Based on these data, calculate the satellite communication blind area distribution: blind area probability calculation method: P(x,y) = P_s·P_m·P_t; where P(x,y) is the blind area probability of the coordinate point (x,y); P_s is the terrain occlusion probability, which is calculated by the ray tracing method to count the number of intersection points of the terrain and the satellite connection line. In this embodiment, P_s = 0.85 in mountain area A; P_m is the probability affected by meteorological factors, P_m = 1 - exp(-α·R), where α is the meteorological attenuation coefficient 0.042, and P_m = 0.30 is calculated; P_t is the time-related factor, and P_t = 0.92 is calculated according to the satellite orbit parameters. Substitute the formula to calculate the blind area probability of area A, P(A) = 0.85×0.30×0.92 = 0.23.

[0162] The system divides the blind area into a complete blind area (P>0.7) and an edge blind area (0.2<P≤0.7). In this embodiment, by counting the blind area probabilities of 5000 grid points, 4 complete blind area regions and 12 edge blind area regions are identified.

[0163] S12. Adaptive multimodal image acquisition parameter configuration.

[0164] For the identified blind areas, the system configures image acquisition parameters: resolution selection function: R(d) = R_max·(1-β·d); where R is the acquisition resolution; R_max is the maximum resolution parameter of 4096×3072 pixels; d is the normalized distance to the edge of the blind area (between 0 and 1); β is the resolution attenuation coefficient of 0.6. For example, for the area d=0.25, R=4096×(1-0.6×0.25)=3686 pixels is calculated.

[0165] Scene complexity score calculation: SC = w_1·E + w_2·V + w_3·C; where SC is the scene complexity score; E is the image entropy value (7.8 for sampling point A); V is the region variance (32.5 for sampling point A); C is the edge density (0.45 for sampling point A); w_1, w_2, and w_3 are weighting coefficients, respectively, of 0.5, 0.3, and 0.2. Substituting into the formula, the SC for sampling point A is calculated as 0.5×7.8+0.3×32.5+0.2×0.45=13.69, which exceeds the preset threshold of 7.5, so the system automatically switches to multispectral acquisition mode.

[0166] Using the above configuration parameters, the UAV performed multimodal image acquisition, capturing 458 raw images. Each image was appended with metadata tags, including GPS coordinates, timestamp, acquisition mode, and quality metrics, forming a labeled image dataset.

[0167] S2. Establish a low-altitude UAV relay network based on labeled image datasets.

[0168] Construct a relay network and implement link monitoring and maintenance.

[0169] S21. Relay network topology initialization.

[0170] Based on the labeled image dataset, the positions of four UAVs were initialized: UAV1 (34.05°N, 108.92°E, altitude 150m), UAV2 (34.07°N, 108.93°E, altitude 180m), UAV3 (34.08°N, 108.94°E, altitude 220m), and UAV4 (34.06°N, 108.95°E, altitude 200m).

[0171] The system performs network connectivity tests, measuring the initial communication quality indicators of each link: Link L12 (UAV1→UAV2): Signal-to-noise ratio (SNR) = 18.5 dB, Bit error rate (BER) = 2.3 × 10⁻⁶ -5 Link L23 (UAV2→UAV3): Signal-to-noise ratio (SNR) = 16.8 dB, Bit error rate (BER) = 4.1 × 10⁻⁶ -5Link L34 (UAV3→UAV4): Signal-to-noise ratio (SNR) = 15.2 dB, Bit error rate (BER) = 5.7 × 10⁻⁶ -5 Link L41 (UAV4→UAV1): Signal-to-noise ratio (SNR) = 14.3 dB, Bit error rate (BER) = 8.2 × 10⁻⁶ -5 .

[0172] Analyzing the test results, a network vulnerability map was generated, identifying link L41 as the most vulnerable link (overall score 0.67). Applying a topology self-optimization algorithm, the position of UAV4 was adjusted to (34.063°N, 108.948°E), and the antenna pointing angle was adjusted by 15°, improving the signal-to-noise ratio of link L41 to 17.8dB and reducing the bit error rate to 3.2×10⁻⁶. -5 The overall score improved to 0.82.

[0173] S22. Implementation of predictive link degradation detection.

[0174] The system continuously monitors the communication quality of each relay node and updates the communication quality heatmap in real time. For link L23, the system collects the following time-series data (5 sampling points within the past 10 minutes): t1: SNR=16.8dB, BER=4.1×10 -5 t2: SNR=16.5dB, BER=4.6×10 -5 ;t3: SNR=16.1dB, BER=5.3×10 -5 ;t4: SNR=15.7dB, BER=5.9×10 -5 ;t5: SNR=15.2dB, BER=6.8×10 -5 ;

[0175] Based on this time-series data, the system performs link health early warning calculations: Link stability feature extraction: The system calculates statistical features to obtain the basic statistical feature set F_s={μ_SNR=16.06, σ 2 _SNR=0.32, μ_BER=5.34×10 -5 , σ 2 _BER=9.12×10 -11 Frequency domain features are extracted using wavelet transform, resulting in a frequency domain feature vector F_f = {0.42, 0.21, 0.08}. Trend features are calculated, resulting in a trend feature vector F_t = {Δ_SNR = -0.4, Δ...}. 2 _SNR=-0.05, S_BER=6.75×10 -6}, where Δ represents the first-order difference, Δ 2 Let S represent the second difference, and let S represent the slope of the linear regression.

[0176] Degradation pattern matching: Signal attenuation pattern M_a, interference surge pattern M_i, and multipath effect enhancement pattern M_m are extracted from a predefined degradation pattern library. The similarity between the feature set and each pattern is calculated: S(F,M_a) = DTW(F,M_a) / (1+λ·L) = 0.13 / (1+0.05×10) = 0.087; S(F,M_i) = 0.21 / (1+0.05×10) = 0.14; S(F,M_m) = 0.17 / (1+0.05×10) = 0.113; where DTW is the dynamic time warping distance, L is the sequence length (10), and λ is the length penalty coefficient (0.05). The highest similarity is for the interference surge pattern (0.14), and the system determines the current degradation pattern to be interference surge.

[0177] Degradation trend prediction based on residual network: The system inputs the feature matrix to a pre-trained residual network to predict communication quality indicators at future time points: t+1min: SNR*=14.8dB, BER*=7.5×10 -5 ;t+3min: SNR*=14.2dB, BER*=8.6×10 -5 ;t+5min: SNR*=13.6dB, BER*=9.9×10 -5 The prediction formula is: Q*(t+Δt) = Q(t) + ΔQ(t,Δt); Q*(t+Δt) is the predicted communication quality index at the future time point; Q(t) is the communication quality index at the current time point; ΔQ is the change in the residual network prediction. Link health warning is generated: the system calculates the degradation risk coefficient at each time point (link reliability decreases significantly when SNR < 14dB): t+1min: risk = 0.34 (slight); t+3min: risk = 0.65 (moderate); t+5min: risk = 0.82 (severe). The system generates a warning message: "Link L23 will degrade to an unstable state in approximately 4 minutes. It is recommended to adjust the UAV3 position or increase power, confidence level 78%." Based on this warning, the system triggers a minor adjustment of the UAV's position, moving UAV3 25 meters northeast, thus avoiding the risk of link interruption in advance.

[0178] S3. Perform feature extraction and semantic analysis on the labeled image dataset to generate non-uniform image slices and transmission priorities.

[0179] S31. Image semantic analysis and segmentation map generation.

[0180] The system performs semantic analysis on the labeled image dataset. Taking a representative image (3686×2765 pixels) as an example, a multi-resolution image pyramid is first constructed to generate a 4-level scale representation. A lightweight disaster feature extractor is applied to extract basic feature vectors, including: edge features (Sobel operator): v_edge = {0.42, 0.58, 0.31, 0.65}; texture features (LBP): v_texture = {0.28, 0.63, 0.52, 0.39, 0.47}; color features (HSV histogram): v_color = {0.35, 0.42, 0.28, 0.56, 0.63}; and shape features (Hu moments): v_shape = {0.25, 0.31, 0.42, 0.29}. These feature vectors are then input into a disaster scene classification model to obtain the scene type probability distribution: building damage: p=0.62; road interruption: p=0.45; crowd gathering: p=0.38; Fire area: p=0.12; Flooded area: p=0.08; Based on the recognition results, the system performs semantic segmentation, generates a semantic segmentation map, and marks 7 key areas: Personnel (No. 01): Location (1250, 980), area 1.2%, importance score 0.95; Severely damaged building (No. 02): Location (820, 650), area 8.5%, importance score 0.88; Slightly damaged building (No. 03): Location (1500, 1200), area 12.3%, importance score 0.65; Road interruption (No. 04): Location (950, 1400), area 4.8%, importance score 0.82; Evacuation route (No. 05): Location (1850, 950), area 5.2%, importance score 0.85; Open space (No. 06): Location (2200, 1800): Area 25.8%, Importance Score 0.42; Background Area (No. 07): Area 42.2%, Importance Score 0.25.

[0181] S32. Semantic-aware non-uniform image segmentation generation.

[0182] The system extracts semantic boundaries and importance features based on a semantic segmentation map to generate a segmentation guidance map. The system applies the Canny edge detection algorithm (low threshold 0.15, high threshold 0.35) to process the semantic segmentation map, enhancing semantic region boundaries and generating an enhanced boundary map E_b. Importance weights are assigned to semantic regions: personnel area: w=0.95; severely damaged buildings: w=0.88; road interruption: w=0.82; evacuation routes: w=0.85; slightly damaged buildings: w=0.65; open space: w=0.42; background area: w=0.25. The system calculates the texture complexity of each region using a complexity map C_t calculated from local image entropy (8×8 window). For example, the average complexity for personnel areas is 0.78, and the average complexity for background areas is 0.31. The enhanced boundary map E_b, the importance weight map W_i, and the complexity map C_t are integrated, and a multi-layer weighted algorithm is applied: G(x,y) = 0.4·E_b(x,y) + 0.4·W_i(x,y) + 0.2·C_t(x,y); where G(x,y) is the segmentation guidance map value at coordinates (x,y); E_b(x,y) is the enhanced boundary map value; W_i(x,y) is the importance weight map value; C_t(x,y) is the complexity map value; and 0.4 and 0.2 are weight coefficients.

[0183] An adaptive quadtree segmentation algorithm is applied: The system performs initial quadtree segmentation on the image based on the segmentation guidance map G value, dividing the image into blocks of varying sizes. Segmentation rules: If the variance of the G value within a region is greater than a threshold (T=0.15) and the region is larger than the minimum size (32×32 pixels), segmentation continues; otherwise, segmentation stops. The quadtree segmentation structure is evaluated, identifying the following problems: Semantic fragmentation: The personnel area is divided into 4 different segments (P1, P2, P3, P4); Boundary misalignment: The boundary of the road interruption area deviates from the segment boundary by up to 15 pixels; Scale imbalance: The segment sizes of the evacuation passage area vary greatly (maximum 512×512, minimum 32×32); Island effect: The personnel area forms a small 32×32 segment, creating islands with the surrounding 128×128 segments. Semantic boundary-driven boundary adjustment: To address the boundary misalignment problem, the system applies a semantic boundary adjustment algorithm: For identified misaligned regions, the segment boundary is moved towards the nearest semantic boundary, with a maximum movement distance of 15% of the original boundary length. For road interruption areas, the original segment boundary coordinates (950, 1385) are adjusted to (950, 1400), perfectly aligned with the semantic boundary, generating a boundary-optimized segment structure.

[0184] The system performs semantic integrity guarantee processing on the boundary-optimized shard structure: overlay the boundary-optimized shard structure with the semantic segmentation map, and calculate the semantic object segmentation situation: Personnel area (number 01): Distributed in 4 shards, the largest shard accounts for 45%, exceeding the preset threshold of 30%, marked as a semantically fragmented area; Calculate the semantic fragmentation loss: L1 = w×(p - t) = 2.5×(0.45 - 0.3) = 0.375; where w is the weight coefficient 2.5; p is the maximum segmentation ratio 0.45; t is the preset threshold 0.3.

[0185] Construct the shard merging cost function: The cost C12 of merging shards P1 and P2 is 0.28 (based on the boundary importance score); The cost C13 of merging shards P1 and P3 is 0.32; The cost C24 of merging shards P2 and P4 is 0.25; Since C12 < L1 (0.28 < 0.375), the system performs the merging of P1 and P2 to form a new shard P12. Since C24 < L1 (0.25 < 0.375), the system performs the merging of P2 and P4 to form a new shard P24. For the merged super-large shard P12 (512×384 pixels), the system divides it along the path with the minimum importance gradient, that is, the path avoiding the personnel position, to generate two shards P12a (288×384) and P12b (224×384). Apply the adaptive scale balance algorithm to handle the scale imbalance problem: Calculate the shard size variation coefficient: CV = σ / μ = 112 / 196 = 0.57; The number of shards in each size interval: [32×32]: 18, [64×64]: 32, [128×128]: 25, [256×256]: 10, [512×512]: 4. For high-value areas (personnel and severely damaged buildings), adjust the shard size according to the transmission window characteristics (estimated 30 seconds available): The shard size of the personnel area is adjusted to 64×64 pixels to ensure transmission within a short window; The shard of the severely damaged building is adjusted to 128×128 pixels; Finally, 85 non-uniform image shards are generated, with sizes ranging from 32×32 to 512×512, forming a non-uniform image shard set adapted to the characteristics of the disaster scene.

[0186] S33. Transmission priority calculation.

[0187] Patch base priority calculation: Calculate the base priority for each non-uniform image patch. Taking three representative patches as an example: Patch A (containing personnel area): Disaster type weight T_A = 0.95 (personnel area); Element importance I_A = 0.95; Area proportion A_A = 0.012; P_base(A) = 0.5×0.95 + 0.3×0.95 + 0.2×0.012 = 0.477 + 0.285 + 0.0024 = 0.7644;

[0188] Segment B (containing severely damaged buildings): Disaster type weight T_B = 0.88; Element importance I_B = 0.88; Area proportion A_B = 0.085; P_base(B) = 0.5×0.88 + 0.3×0.88 + 0.2×0.085 = 0.44 +0.264 + 0.017 = 0.721;

[0189] Segment C (including road interruption): Disaster type weight T_C = 0.82; Element importance I_C = 0.82; Area proportion A_C = 0.048; P_base(C) = 0.5×0.82 + 0.3×0.82 + 0.2×0.048 = 0.41 + 0.246 + 0.0096 = 0.6656.

[0190] Read the semantic segmentation map and identify the segment pairs with spatial adjacency and semantic continuity: segment A and segment D (starting point of personnel evacuation route); segment C and segment E (connection point between road interruption and evacuation route). Calculate the cosine similarity of semantic feature vectors: Semantic feature vector V_A = {0.95, 0.42, 0.18, 0.75} for segment A; semantic feature vector V_D = {0.85, 0.38, 0.22, 0.70} for segment D; cosine similarity cos(V_A, V_D) = (V_A·V_D) / (||V_A||·||V_D||) = 0.86; semantic feature vector V_C = {0.82, 0.65, 0.35, 0.42} for segment C; semantic feature vector V_E = {0.85, 0.58, 0.32, 0.48} for segment E; cosine similarity cos(V_C, V_E) = 0.92; generate the association strength matrix, with some values ​​as follows: C(A,D) = 0.86 (exceeds threshold 0.75); C(C,E) = 0.92 (exceeds threshold 0.75); C(A,B) = 0.65 (below threshold 0.75).

[0191] Based on the association strength matrix, fragments with an association strength exceeding the threshold of 0.75 are organized into priority groups: Group G1: {A, D}, group transmission sequence number G(1) = max(P_base(A), P_base(D)) = max(0.7644, 0.682) = 0.7644; Group G2: {C, E}, group transmission sequence number G(2) = max(P_base(C), P_base(E)) = max(0.6656, 0.693) = 0.693. Calculate the relative priority within the group: R(A,1) = P_base(A) / sum(P_base(j)), j∈G1 = 0.7644 / (0.7644+0.682) = 0.529; R(D,1) = 0.682 / (0.7644+0.682) = 0.471; R(C,2) = 0.6656 / (0.6656+0.693) = 0.490; R(E,2) = 0.693 / (0.6656+0.693) = 0.510; Final transmission priority: P(A) = G(1) + σ·R(A,1) = 0.7644 + 0.1×0.529 = 0.81734; P(D) = G(1) + σ·R(D,1) = 0.7644 + 0.1 × 0.471 = 0.81114; P(C) = G(2) + σ·R(C,2) = 0.693 + 0.1 × 0.490 = 0.742; P(E) = G(2) + σ·R(E,2) = 0.693 + 0.1 × 0.510 = 0.744; where σ is the intra-group adjustment coefficient of 0.1. Calculate the transmission priority of all 85 fragments to form a complete transmission priority sequence.

[0192] S4. Implement dual-loop feedback transmission scheduling.

[0193] S41. Generation of fine-grained scheduling plans from three-dimensional optimization models.

[0194] The system employs a three-dimensional optimization model of time, value, and size to generate a refined scheduling plan. A three-dimensional parameter space is constructed: Transmission priority sequences and communication environment states are read to generate a fragmentation basic parameter table (partial data): Fragment ID, Priority, Size (KB), Estimated Transmission Time (s); A 0.81712 0.8; B 0.72148 3.2; C 0.74232 2.1; D 0.81116 1.1; E 0.74428 1.9. Transmission window characteristics are extracted from the comprehensive resource state diagram: Estimated start time: t0 = 14:25:18; Estimated duration: Δt = 30 seconds; Bandwidth fluctuation pattern: Initial 15Kbps, decreasing to 12Kbps in the middle stage, and fluctuating between 8-10Kbps in the late stage.

[0195] The transmission window is divided into three phases: early phase (0-10 seconds): weight configuration W(1) = {wV=0.5, wT=0.2, wS=0.3}; middle phase (10-20 seconds): weight configuration W(2) = {wV=0.4, wT=0.4, wS=0.2}; late phase (20-30 seconds): weight configuration W(3) = {wV=0.6, wT=0.3, wS=0.1}. Calculate the comprehensive value score for segment A: Basic value score Vb(A) = 0.817; Correlation adjustment value Vr(A) = 0.15 (strong correlation with segment D); Timeliness coefficient Ve(A) = 0.2 (high timeliness of personnel location); Comprehensive value score V(A) = Vb(A) + α·Vr(A) + β·Ve(A) = 0.817 + 0.3×0.15 + 0.4×0.2 = 0.862; where α is the correlation value coefficient of 0.3; β is the timeliness coefficient of 0.4; Calculate the comprehensive value scores for other segments similarly: V(B) = 0.735, V(C) = 0.775, V(D) = 0.841, V(E) = 0.778.

[0196] Transmission Risk Assessment: The system uses the Monte Carlo method to perform 1000 simulations, generating a time risk matrix (partial values): Fragment ID t=5st=15st=25s; A 0.05 0.15 0.35; B 0.20 0.40 0.75; C 0.15 0.35 0.65; D 0.08 0.20 0.45; E 0.12 0.30 0.60. Risk coefficient calculation formula: R(i,t) = 1 - Pr(Tc(i) < Tw(t)); where R(i,t) is the transmission risk coefficient of fragment i at time t; Tc(i) is the random variable of fragment transmission completion time; Tw(t) is the random variable of the remaining time of the transmission window; Pr is the probability calculation function.

[0197] Size adjustment strategy generation: Based on the current bandwidth and fragment size, calculate the transmission efficiency coefficient: Early stage (15Kbps): Large fragment size (>32KB) efficiency coefficient 1.2, small fragment size efficiency coefficient 0.8; Mid-stage (12Kbps): Large fragment size efficiency coefficient 1.0, small fragment size efficiency coefficient 1.0; Late stage (8-10Kbps): Large fragment size efficiency coefficient 0.7, small fragment size efficiency coefficient 1.3. Generate phased size adjustment strategies: Early stage: Reduce large size penalty (×0.6) to improve bandwidth utilization; Mid-stage: Balanced processing (×1.0); Late stage: Increase large size penalty (×1.5), prioritize small-volume, high-value fragments.

[0198] Construct the optimization objective function: F(i,t) = wV(t)·V(i) - wT(t)·R(i,t) - wS(t)·S(i)·K(t); where F(i,t) is the optimization objective function value of partition i at time point t; wV(t), wT(t), and wS(t) are the three-dimensional weights corresponding to time point t; V(i) is the comprehensive value score of partition i; R(i,t) is the time risk coefficient; S(i) is the partition size adjustment coefficient; and K(t) is the stage size penalty coefficient. Calculate the objective function values ​​for the early stage (t=5s): F(A,5) = 0.5×0.862 - 0.2×0.05 - 0.3×(12 / 128)×0.6 = 0.431 - 0.01 - 0.0169 = 0.4031; F(B,5) = 0.5×0.735 - 0.2×0.20 - 0.3×(48 / 128)×0.6 = 0.3675 - 0.04 - 0.0675 = 0.26. Calculate the objective function values ​​for the late stage (t=25s): F(A,25) = 0.6×0.862 - 0.3×0.35 - 0.1×(12 / 128)×1.5 = 0.5172 - 0.105 - 0.0141 = 0.3981; F(B,25) = 0.6×0.735 - 0.3×0.75 - 0.1×(48 / 128)×1.5 = 0.441 - 0.225 - 0.0563 = 0.1597. The fine-grained scheduling plan is obtained by solving the problem using the particle swarm optimization algorithm (50 particles, 200 iterations): Early stage transmission order: B→C→E→A→D (prioritizing large-volume, high-value fragments); Mid-stage transmission order: A→D→E→C (balancing value and risk); Late stage transmission order: A→D→F→G (prioritizing small-volume, high-value fragments, with F and G being other small fragments).

[0199] S42. Generation of transmission control commands.

[0200] Constructing a real-time communication state model: The system collects real-time communication state data at 50-millisecond intervals (t=10.2s): Original parameters: SNR=16.3dB, BER=5.2×10 -5 Bandwidth = 11.8Kbps, delay = 185ms. Kalman filtering is applied to eliminate noise: State prediction equation: x_ k- = Ax_ k-1 + Bu_k; where x_ k- Let x_k be the prior state estimate at time k; A is the state transition matrix; x_k is the state estimate at time k. k-1 Here, B is the posterior state estimate at time k-1; B is the control input matrix; u_k is the control vector; the measurement update equation is: x_k = x_k k- + K_k(z_k - Hx_ k- ); where x_k is the posterior state estimate at time k; K_k is the Kalman gain at time k; z_k is the measurement at time k; H is the measurement matrix; Kalman gain calculation: K_k = P_k - H T (HP_k - H T + R) -1 ;where P_k - R is the prior estimation error covariance; R is the measurement noise covariance; applied parameters: state transition matrix A=[1.0, 0.05; 0, 1.0], measurement matrix H=[1.0, 0], process noise covariance Q=diag([0.01, 0.1]), measurement noise covariance R=0.1; filtered communication parameters: SNR=16.2dB, BER=5.0×10 -5 Bandwidth = 12.0Kbps, latency = 180ms; predicted short-term trend (future 200ms): SNR change rate -0.3dB / s, BER change rate +0.8×10 -5 / s, bandwidth change rate -0.5Kbps / s.

[0201] Parameter adjustment and smoothing adjustment calculations: Based on the real-time communication state model, a matching strategy is queried from the transmission parameter adjustment rule base: when the bandwidth decrease rate > 0.3 Kbps / s, the compression ratio is increased; when the BER increase rate > 0.5 × 10 -5 At a rate of / s, increase the forward error correction level. Apply a fuzzy control algorithm to determine the adjustment intensity: the bandwidth decrease rate (-0.5Kbps / s) belongs to the "medium" level (membership 0.75); the BER increase rate (+0.8×10) -5 / s) belongs to the "faster" level (membership 0.85); Generation parameter adjustment: Compression ratio adjustment: Δcr = +10% (increase compression); Forward error correction level adjustment: Δfec = +1 level; Packet size adjustment: Δps = -128 bytes.

[0202] The system generates smoothing adjustments: It reads parameter adjustments and historical adjustment records (past 3 periods): t-3: Δcr=-5%, Δfec=0, Δps=-64; t-2: Δcr=+8%, Δfec=+1, Δps=-128; t-1: Δcr=+12%, Δfec=0, Δps=-128. It calculates adjustment trend indicators: Compression ratio direction change frequency: 1 / 2 (one direction change divided by 2 intervals); Forward error correction level direction change frequency: 2 / 2; Packet size direction change frequency: 0 / 2.

[0203] Identifying parameter oscillation risks: Compression ratio oscillation risk score Osc(cr) = 1 / 2 = 0.5; Forward error correction level oscillation risk score Osc(fec) = 2 / 2 = 1.0; Packet size oscillation risk score Osc(ps) = 0 / 2 = 0. Dynamically constructing the adaptive damping function: D(p) = D_min + (D_max - D_min)·Osc(p) η Where D(p) is the damping coefficient of parameter p; D_min is the minimum damping coefficient (0.2); D_max is the maximum damping coefficient (0.8); Osc(p) is the oscillation risk score of parameter p; and η is the nonlinear adjustment index (2.0). The damping coefficients for each parameter are calculated as follows: D(cr) = 0.2 + (0.8 - 0.2) × 0.5 2 = 0.2 +0.6×0.25 = 0.35; D(fec) = 0.2 + (0.8-0.2)×1.0 2 = 0.2 + 0.6×1.0 = 0.8; D(ps)= 0.2 + (0.8-0.2)×0 2 = 0.2. Generate smoothing adjustment amount: Δcr' = Δcr×(1-D(cr)) = 10%×(1-0.35) = 6.5%; Δfec' = Δfec×(1-D(fec)) = 1×(1-0.8) = 0.2 -> 0 (rounded); Δps' = Δps×(1-D(ps)) = -128×(1-0.2) = -102.4 -> -100 bytes (rounded).

[0204] Transmission control command generation: Apply smoothing adjustments to current transmission parameters: Current compression rate: 60% -> New compression rate: 66.5%; Current forward error correction level: Level 2 -> New level: Level 2 (remains unchanged); Current packet size: 1024 bytes -> New packet size: 924 bytes; Generate updated parameter values ​​and anomaly response commands: Parameter update: {cr:66.5%, fec:2,ps:924, timeout:2.5s}; Anomaly response: "Increased interference detected, enable frequency hopping mode, prepare backup channel"; Generate transmission control command: "Parameter update: cr=66.5%, fec=2,ps=924, timeout=2.5s; Response: Enable frequency hopping mode, prepare backup channel".

[0205] S43. Fine-grained scheduling plan update.

[0206] Comprehensive performance evaluation: The system collects transmission confirmation information every 500 milliseconds (t=15.5s): Completed fragments: B, C, E, A (partial); Transmission success rate: 87.5%; Actual transmission rate: 11.3Kbps (5.8% lower than the expected 12Kbps); Resource utilization efficiency: 78%. Reading the fast loop operation logs from the past 10 seconds and analyzing the effects: Average effectiveness of compression ratio adjustment: 82%; Average effectiveness of packet size adjustment: 75%; Average effectiveness of forward error correction level adjustment: 90%. Status of fragments awaiting transmission in the queue: Fragment A: 68% transmitted, 3.84KB remaining; Fragment D: Not started, 16KB; Fragment F: Not started, 8KB; Fragment G: Not started, 6KB. Comprehensive performance evaluation: "The transmission rate in the mid-stage is 5.8% lower than expected, but the success rate is up to standard (87.5%). Fragment A is 68% complete, expected to take an additional 0.7 seconds to complete. The adjustment strategy is effective, and the resource utilization rate is 78%."

[0207] Transmission window status update and phase switching. Based on the transmission rate data in the comprehensive performance evaluation, the transmission window model is updated: original window remaining time: 15 seconds; rate decrease rate: 5.8%; estimated remaining time after update: 15 × (1 - 0.058) = 14.13 seconds; window reliability: updated from 87% to 80%; window status update is generated: "Transmission window remaining time updated to 14.13 seconds, reliability decreased to 80%, expected window end time 14:25:48". The current transmission window phase is detected (15.5 seconds have elapsed, total time 30 seconds): original plan: mid-stage (10-20 seconds); based on the rate decrease and remaining time, it is necessary to switch to the late-stage strategy in advance; phase switching signal is generated: "Switch to the late-stage in advance, time point adjusted to t=16 seconds". Analysis of transmission bottlenecks: Primary bottleneck: Increased packet retransmission rate (5.8%, exceeding the threshold of 3%); Secondary bottleneck: Increased bandwidth fluctuation (standard deviation of 1.2Kbps, exceeding the threshold of 0.8Kbps); Performance bottleneck report generated: "Primary bottleneck: Packet retransmission rate 5.8% (exceeding the standard), it is recommended to increase the forward error correction level; Secondary bottleneck: Increased bandwidth fluctuation, it is recommended to reduce the packet size and enable bandwidth prediction."

[0208] Based on window status updates, phase switching signals, and performance bottleneck reports, a transmission strategy adjustment plan is formulated: immediately switch to a late-stage transmission strategy; increase the forward error correction level to level 3 (covering major bottlenecks); further reduce the packet size to 768 bytes (addressing minor bottlenecks); adjust fragmentation priorities, prioritizing the completion of high-value, small-volume fragment transmission; generate a strategy adjustment recommendation: "Switch to a late-stage strategy, increase the forward error correction level to level 3, adjust the packet size to 768 bytes, prioritize the transmission of the remaining part of fragment A, D, F, and G, and abandon low-value, large-volume fragments." Based on the strategy adjustment recommendation, update the scheduling plan: update the fragmentation order: A (remaining part) → D → F → G → H → ...; update the time allocation: A (remaining) allocated 0.7 seconds, D allocated 2.5 seconds, F allocated 1.3 seconds, and G allocated 1.0 second. Adjust resource allocation and encoding strategies: apply a higher compression rate (70%) to fragments D, F, and G; upgrade the forward error correction level to level 3; enable bandwidth prediction. The updated fine-grained scheduling plan is as follows: "Window phase: late; priority transmission: A (remaining, 0.7 seconds), D (2.5 seconds), F (1.3 seconds), G (1.0 seconds); parameter configuration: compression rate 70%, forward error correction level 3, packet size 768 bytes; special measures: enable bandwidth prediction, conditionally abandon low-value fragments."

[0209] S5. Evaluation of actual transmission performance.

[0210] Transmission window usage: Actual window duration: 28.5 seconds (5% shorter than the expected 30 seconds); Effective bandwidth utilization: average 79.2%; Number of successfully transmitted fragments: 56 (out of a total of 85, success rate 65.9%); Value transmission efficiency: High-value fragment (priority > 0.7) transmission success rate: 92.3% (24 / 26); Medium-value fragment (0.5-0.7) transmission success rate: 68.4% (26 / 38); Low-value fragment (< 0.5) transmission success rate: 28.6% (6 / 21); Value-weighted transmission success rate: Σ(P(i)·S(i)) / Σ(P(i)) = 84.5%; where P(i) is the priority of fragment i; S(i) is the transmission success indicator (1 for success, 0 for failure). System response capability assessment: Recovery time to bandwidth surge (35% bandwidth drop at t=18s): 1.2 seconds; Emergency handling of early communication window closure: Successfully switched to high-value priority mode, ensuring the transmission of critical information; Semantic association fragment transmission integrity: 87.5% of associated fragment groups were transmitted completely. Value-weighted transmission success rate of traditional uniform fragmentation method: 53.2% (31.3 percentage points lower); Bandwidth utilization rate of traditional single-cycle feedback method: 62.8% (16.4 percentage points lower); High-value fragment transmission success rate of the windowless transmission awareness method: 71.5% (20.8 percentage points lower). This embodiment fully demonstrates the implementation process of a method for image acquisition and transmission in satellite communication blind spots at disaster sites using low-altitude UAV relay. The dual-loop feedback mechanism achieves a balance between rapid parameter response and stable strategy optimization by separating a 50ms fast loop and a 500ms slow loop, improving the system's adaptability to fluctuating communication environments. The time-value-size three-dimensional optimization model achieves transmission window awareness through dynamic partitioning of the transmission window and phased weight configuration, employing differentiated strategies at different stages to ensure the transmission efficiency of high-value information. Semantic-aware non-uniform segmentation, based on semantic segmentation and an adaptive quadtree algorithm, ensures the integrity of semantic objects and ensures the coherent transmission of relevant information through semantic association priority calculation. Adaptive damping control dynamically adjusts the damping coefficient through oscillation risk scoring, effectively preventing parameter oscillations and improving system stability.

[0211] This invention constructs a separate fast-slow dual-loop feedback architecture, separating real-time parameter adjustment and strategy optimization functions while allowing them to work collaboratively. The fast loop, with a period of 50 milliseconds, collects real-time communication status and responds quickly to communication fluctuations through Kalman filtering and predictive algorithms. The slow loop, with a period of 500 milliseconds, collects transmission confirmation information and adjusts medium- to long-term transmission strategies, resolving the contradiction between response speed and strategy stability that is difficult to achieve in single-cycle feedback systems. The adaptive damping mechanism introduced in the fast loop can automatically identify parameter oscillation risks and dynamically adjust the damping coefficient, avoiding over-adjustment of parameters in environments with severe fluctuations. By analyzing the frequency and amplitude of adjustment direction changes in the most recent three cycles, the system can apply a strong damping coefficient to parameters with high oscillation risk and a weak damping coefficient to parameters with low risk, thereby effectively suppressing parameter oscillations while maintaining rapid response capabilities and improving the overall stability of the system. A groundbreaking three-dimensional optimization model of time-value-size is proposed, incorporating the transmission window time dimension as a core factor into the optimization decision. The system divides the transmission window into three stages: early, middle, and late. The weights of these three dimensions are dynamically adjusted for each stage: the early stage prioritizes bandwidth utilization, focusing on transmitting high-value, large-volume fragments; the middle stage balances the three dimensions; and the late stage increases the weight of the value dimension, prioritizing the transmission of high-value, small-volume fragments. The transmission risk assessment in the model employs Monte Carlo simulation, calculating the transmission risk coefficients of fragments at different time points through numerous random simulations. This allows the system to make more reliable decisions based on probabilistic models rather than simple deterministic estimations. This "wide at the beginning, narrow at the end" transmission strategy is particularly suitable for unstable transmission environments at the edge of satellite communication blind zones, maximizing the transmission of information value within a limited transmission window. A semantically aware adaptive quadtree segmentation method is designed, particularly a two-layer fragmentation strategy, which addresses the unique problem of uneven distribution of information value in disaster images. Based on regional dynamic characteristics, image regions are divided into high-dynamic regions (such as personnel locations and the forefront of fire spread) and low-dynamic regions, applying fine-grained and coarse-grained fragmentation parameters respectively to form a two-layer fragmentation structure. By establishing inter-layer mapping relationships and optimizing boundary consistency, the system ensures that fragmentation does not sever important semantic objects while adapting to different transmission conditions. During transmission, the system dynamically selects which layer of slicing to use based on the transmission window status and regional change monitoring results. This enables precise transmission of high-value, rapidly changing areas and efficient transmission of general areas, cleverly resolving the contradiction between semantic integrity and transmission flexibility that single-scale slicing cannot simultaneously satisfy. Furthermore, this invention introduces a semantic association priority calculation mechanism. By calculating the cosine similarity of semantic feature vectors between slicings, it identifies and organizes groups of slicings with semantic associations, ensuring the coherent transmission of semantically related content and further enhancing the system's ability to protect the semantic information of disaster images.

[0212] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of 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 protection scope of the present invention.

Claims

1. A method for acquiring and transmitting images of disaster sites in satellite communication blind spots via low-altitude unmanned aerial vehicle (UAV) relay, characterized in that, include: Acquire disaster area information and perform adaptive multimodal image acquisition to generate a labeled image dataset; A low-altitude UAV relay network was established based on a labeled image dataset, and the relay network topology was constructed and maintained. Feature extraction and semantic analysis are performed on the labeled image dataset to generate non-uniform image segments, and the transmission priority of each segment is calculated to form a transmission priority sequence. Monitor the communication environment status of the relay network topology, combine the transmission priority sequence, and dynamically adjust the transmission strategy through a double-loop feedback mechanism to perform adaptive transmission of image fragments; The steps involved in performing feature extraction and semantic analysis on the labeled image dataset to generate non-uniform image patches include: Perform semantic analysis on the labeled image dataset to generate a semantic segmentation map; Based on the semantic segmentation map, semantic boundaries and importance features are extracted to generate a segmentation guidance map; Based on the segmentation guidance map, an adaptive quadtree segmentation algorithm is applied to generate the initial fragmentation structure; Perform semantic boundary-driven boundary adjustment on the initial fragmentation structure to generate a boundary-optimized fragmentation structure; Semantic integrity protection processing is performed on the boundary-optimized segmentation structure to generate non-uniform image segments adapted to the characteristics of disaster scenarios.

2. The method according to claim 1, characterized in that, The steps for dynamically adjusting the transmission strategy using a dual-loop feedback mechanism include: Read the transmission priority sequence and communication environment status, apply the time-value-size three-dimensional optimization model, and generate a fine-grained scheduling plan, i.e., a transmission strategy; Real-time communication status is collected at the first preset period, and transmission control commands are generated in combination with a fine scheduling plan. The transmission confirmation information is collected in the second preset cycle to evaluate the execution effect of the transmission control command and update the fine scheduling plan. The first preset cycle is shorter than the second preset cycle, forming a dual-loop feedback structure for rapid parameter adjustment and strategy optimization.

3. The method according to claim 2, characterized in that, The steps for generating a fine-grained scheduling plan by applying a three-dimensional optimization model of time, value, and size include: Read the transmission priority sequence and communication environment status, construct a three-dimensional parameter space, and generate a fragmentation basic parameter table and transmission window characteristic description; based on this, divide the transmission window into three stages: early, middle and late, and set differentiated weight configurations for each stage; Based on the basic parameter table for segmentation, the value of segmented content, the correlation adjustment value, and the timeliness coefficient are calculated to generate a comprehensive value score. Based on the characteristics of the transmission window, a transmission risk assessment is performed, and a time risk matrix is ​​generated. Based on the analysis of the fragmentation basic parameter table, transmission resource efficiency is determined, and a size adjustment strategy is generated. By inputting the comprehensive value score, time risk matrix, and size adjustment strategy into the optimization objective function, a refined scheduling plan can be obtained.

4. The method according to claim 2, characterized in that, The steps for generating transmission control commands include: Real-time communication status is collected, noise is eliminated by Kalman filtering to generate filtered communication parameters, and short-term parameter change trends are predicted based on these parameters to construct a real-time communication status model. Based on the real-time communication state model, the matching strategy is queried from the transmission parameter adjustment rule base to determine the adjustment intensity and generate parameter adjustment amount and smoothing adjustment amount. By combining the fine-grained scheduling plan, the smooth adjustment amount is applied to the current transmission parameters to generate updated parameter values ​​and anomaly response instructions, thus forming transmission control instructions.

5. The method according to claim 4, characterized in that, The steps for generating the smoothing adjustment amount include: Read parameter adjustment amounts and historical adjustment records, calculate the frequency and magnitude of parameter adjustment direction changes within a predetermined period, and generate adjustment trend indicators; Based on the adjustment trend indicator, potential parameter oscillation risks are identified, and oscillation risk scores for each transmission parameter are calculated. Based on the oscillation risk score, an adaptive damping function is dynamically constructed. Oscillation risk scores exceeding the threshold are treated with a strong damping coefficient, while oscillation risk scores below the threshold are treated with a weak damping coefficient, generating a parameterized damping coefficient table. The parameter adjustment amount is weighted and calculated with the corresponding damping coefficient in the parameterized damping coefficient table. The maximum amplitude and frequency of parameter adjustment are limited within a preset time to generate a smooth adjustment amount.

6. The method according to claim 2, characterized in that, The steps to update a fine-grained scheduling plan include: Collect transmission confirmation information and fast loop operation logs, construct a comprehensive performance evaluation, reassess the remaining time and reliability of the transmission window, generate window status updates, detect changes in the transmission window stage to generate stage switching signals, and analyze bottlenecks in the transmission process to generate performance bottleneck reports. Based on window status updates, phase switching signals, and performance bottleneck reports, formulate a transmission strategy adjustment plan and generate strategy adjustment recommendations. Read the current fine-grained scheduling plan, update the scheduling order and time allocation of each segment according to the strategy adjustment suggestions, and generate an updated scheduling plan; By combining the updated scheduling plan and transmission control instructions, the resource allocation and encoding strategies of subsequent fragments are adjusted to generate a resource optimization scheme; this scheme is then integrated with the updated scheduling plan to form an updated fine-grained scheduling plan.

7. The method according to claim 1, characterized in that, Performing semantic integrity protection processing on the boundary-optimized piecewise structure to generate non-uniform image pieces includes the following steps: Perform fragment merging on the same semantic object that is split in the boundary-optimized fragmentation structure, and split the fragments that exceed the preset threshold after merging along the path with the minimum importance gradient to generate a semantically optimized fragmentation structure. Adaptive scale balancing is applied to the semantically optimized fragmentation structure, and isolated small fragments are extracted. A merging strategy is determined based on semantic similarity and importance to generate a balanced fragmentation structure. For high-value information regions in the balanced processing fragmentation structure, the fragment size is adjusted according to the characteristics of the transmission window, so that key information can be transmitted within a short transmission window, generating non-uniform image fragments.

8. The method according to claim 1, characterized in that, The steps for calculating the transmission priority of each fragment include: Calculate the disaster type, element importance, and area proportion of each non-uniform image patch to generate the basic priority of the patch; Read the semantic segmentation map and basic segmentation priority, identify segment pairs with spatial adjacency and semantic continuity, measure the degree of semantic association by calculating the cosine similarity of semantic feature vectors, and generate an association strength matrix; Based on the correlation strength matrix, fragments with correlation strength exceeding a preset threshold are organized into priority groups, and an overall transmission sequence number is assigned to each group to generate a priority group table. Within each priority group, the relative priority within the group is calculated based on the ratio of the basic priority of the fragmentation, and then combined with the overall transmission sequence number of the group to form an associated adjustment priority, i.e., the transmission priority.

9. The method according to claim 7, characterized in that, Performing semantic integrity protection processing on the boundary-optimized fragmentation structure also includes: Spatial overlay analysis is performed on the boundary optimization piecewise structure and the semantic segmentation map to calculate the proportion of each semantic object being segmented into different pieces. When it exceeds a preset threshold, it is marked as a semantic fragmentation region. Semantic fragmentation region labels are generated and semantic fragmentation loss is calculated. For each semantic object in the semantically fragmented region label, a fragmentation merging cost function is constructed by combining the preset fragmentation boundary importance score. When the fragmentation merging cost is lower than the semantic fragmentation loss, fragmentation merging is performed to generate a semantically merged fragmentation structure. Analyze the size distribution of each piece in the semantic merging fragmentation structure, calculate the coefficient of variation of the fragment size and the distribution of the number of fragments in each size interval, and generate a scale distribution map; By combining a pre-defined importance map and scale distribution map, an adaptive scale adjustment function is constructed to generate a scale-balanced piecewise structure.

Images