Generative ai remote sensing image disaster dynamic monitoring system, device and application

Abstract

The application belongs to the technical field of cross technology of artificial intelligence, satellite remote sensing and insurance technology, and particularly relates to a generative AI remote sensing image disaster dynamic monitoring system, equipment and application. By integrating a fluid mechanics model and a building structure dynamics model, a synthetic dataset with physical labels is generated; satellite images, unmanned aerial vehicle video streams and weather station data are aligned by a three-source space-time alignment module, after which the three-source asynchronous input data is synchronized and a semantic segmentation disaster recognition result is output by a band adaptive attention diffusion model, and then a continuous time domain disaster evolution model is constructed by a neural differential equation deduction engine to output a disaster diffusion path and intensity heat map in T hours in the future; in combination with OpenStreetMap building semantic information and regional economic density atlas, a disaster loss interval valuation report is automatically generated. The application constructs a closed loop system of "physical synthesis, dynamic deduction, insurance pricing and temperature drift correction", forming a monitoring and actuarial integrated solution.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of cross technology of artificial intelligence, satellite remote sensing and insurance technology, and specifically relates to a generative AI remote sensing image disaster dynamic monitoring system, equipment and application. BACKGROUND

[0002] The disaster AI system is a comprehensive monitoring platform integrating artificial intelligence, satellite remote sensing and insurance actuarial technology, aiming to realize real-time perception and rapid response of disaster risk. With the deployment of high-resolution satellite constellation and the improvement of AI computing power, the system gradually replaces the traditional manual interpretation mode in disaster monitoring such as floods and forest fires. However, due to the lack of physical laws, the lack of economic dynamic coupling and the bottleneck of hardware cooperation in the prior art, the monitoring results are difficult to directly drive the insurance actuarial process, which restricts its core role in catastrophic risk management.

[0003] In the traditional technology: ① Lack of physical laws: pure data-driven generative models do not introduce fluid mechanics equations and material strength criteria, resulting in deviation of storm runoff prediction and building damage assessment from natural laws, and significantly reducing the reliability of extreme event prediction; ② Economic dynamics are disconnected: static economic parameters (such as fixed asset density) cannot adapt to regional GDP growth and industrial upgrading, causing systematic deviation of insurance claims and actual losses in high development areas, inducing market risks such as adverse selection; ③ Difficulty in multi-source cooperation: the temperature drift effect of satellite-borne sensors leads to distortion of infrared temperature measurement, PMS / IRS time phase misalignment causes fusion error, and the lack of ground computing power causes disaster deduction time lagging behind the emergency decision-making window.

[0004] Therefore, how to provide an integrated disaster monitoring and insurance pricing method that integrates physical constraint generative AI, dynamic deduction engine, economic dynamic coupling model and hardware temperature drift suppression mechanism is the key to solving the above problems. SUMMARY

[0005] To solve the problems in the prior art, the present application provides a generative AI remote sensing image disaster dynamic monitoring system, equipment and application.

[0006] The application solves the technical problems by adopting the technical scheme as follows: a remote sensing image disaster dynamic monitoring system of generative AI, comprising: a data generation layer, which generates a synthetic dataset with physical labels by integrating a fluid mechanics model and a building structure dynamics model; an intelligent analysis layer, which aligns satellite images, unmanned aerial vehicle video streams and weather station data by using a three-source space-time alignment module, outputs a semantic segmentation disaster recognition result after the synchronous three-source asynchronous input data is output by a band self-adaptive attention diffusion model, and then constructs a continuous time domain disaster evolution model by using a neural differential equation deduction engine to output a disaster diffusion path and intensity heat map in T hours in the future; and an application layer, which automatically generates a disaster loss interval valuation report in combination with OpenStreetMap building semantic information and regional economic density atlas.

[0007] Preferably, the fluid mechanics model simulates fluid motion under different terrains by using a shallow water equation, superimposes a rainfall intensity parameter to generate water body turbidity characteristics, inputs terrain DEM data to generate synthetic flood images, and adopts a variable terrain Manning coefficient calculation method in the fluid simulation process: wherein n represents the variable terrain Manning coefficient, n_0 represents the initial value, and represents the terrain gradient. The building structure dynamics model calculates building pressure deformation based on material mechanics parameters to generate texture details and shadow characteristics of collapsed ruins.

[0008] Preferably, the band self-adaptive attention diffusion model comprises: a near-infrared band feature enhancement layer, which enhances the near-infrared band features by amplifying the band abnormal features through learnable weights: wherein represents the near-infrared band feature enhancement value, and represents the near-infrared band feature. wherein represents the visible light band feature, and W_Q represents a query vector weight matrix and W_K represents a key vector weight matrix. wherein represents an L2 norm operator, d_k represents a key vector dimension, and softmax represents a normalization function.A thermal infrared band noise suppression mechanism: the satellite thermal radiation value offset is corrected by using weather station temperature data; a thermal infrared band drift correction unit: T_adj = T_raw + 0.15·(T_amb - 25); wherein T_adj represents the corrected thermal infrared temperature, T_raw represents the sensor original output value, and T_amb represents the FPGA chip environment temperature.

[0009] ​Preferably, the neural differential equation derivation engine is implemented by constructing an ordinary differential equation for disaster diffusion: In the formula, h represents the hidden state vector of disaster intensity; The parameters are learnable; t represents the disaster evolution time; a sudden event sensing unit is introduced: when the satellite detects that the instantaneous rainfall is greater than the threshold, it automatically switches to a higher-order differential equation. In the formula, I_rain represents the real-time rainfall intensity; γ represents the disaster response intensity caused by unit rainfall acceleration; λ represents the system inertia; h represents the hidden state vector of disaster intensity; t represents the disaster evolution time; the simulation result visualization component generates a three-dimensional diffusion animation with confidence intervals, and supports backtracking the state at any time point.

[0010] Preferably, the disaster loss assessment module in the application layer includes:

[0011] Building type classifier: Receives GDP growth rate ΔGDP and OpenStreetMap building vectors, and identifies residential / factory / farmland based on DeepLabv3+;

[0012] Asset value update engine: Accesses GDP data by industry from the National Bureau of Statistics, assigns loss weights by building type, and dynamically updates the economic density factor ρk. [1 + αk·ΔGDP] ;where, Indicates the dynamic economic density factor; ΔGDP represents the baseline economic density factor; ΔGDP represents the relative rate of change of GDP; k represents the building type index; t represents the evolution time; αk is the transmission coefficient, which varies for different building types: 0.25 for factories, 0.20 for residential buildings, and 0.18 for commercial buildings.

[0013] Output layer: Outputs a dynamic disaster loss heat map. When ρk reaches a certain value, it triggers an actuarial signal, outputs a claims recommendation, and automatically triggers reinsurance agreement terms.

[0014] A disaster monitoring edge computing device, which integrates the aforementioned generative AI-based remote sensing image disaster dynamic monitoring system, includes:

[0015] Onboard FPGA chip: Deploys a lightweight band-adaptive attention diffusion model to support real-time on-orbit processing;

[0016] Drone swarm linkage interface: When the simulation engine predicts that a disaster has spread to the threshold area, the drone swarm is automatically awakened;

[0017] Distributed evidence storage unit: Stores disaster evolution keyframes and loss assessment results on the blockchain.

[0018] The application of the remote sensing image disaster dynamic monitoring system of the generated AI above is applied to a disaster insurance claim service scene.

[0019] Preferably, a "physical synthesis-dynamic deduction-insurance pricing-wind drift correction" closed loop scheme is formed.

[0020] The insurance pricing method comprises the following steps:

[0021] Receiving geographic fence information of a user's insured area;

[0022] When the system predicts that the disaster probability of the area in the next 24 hours is greater than P%, the premium rate is automatically increased, and the premium floating formula is: wherein, represents the dynamic premium; represents the basic premium rate; represents the threshold disaster probability; h represents the disaster intensity; represents a safety threshold, such as a critical value of the disaster intensity that a building can withstand;

[0023] Pushing a loss assessment report within T hours after the disaster occurs.

[0024] Compared with the prior art, the application has the following beneficial effects:

[0025] 1. In the remote sensing image disaster dynamic monitoring system of the generated AI, the shallow water equation and the material yield strength are used as forced constraint conditions of the generated AI, so that the synthesized disaster data strictly follow the laws of fluid mechanics and structural mechanics, the rainstorm prediction accuracy is significantly improved, the building collapse misjudgment rate is greatly reduced, and the technical defects of the traditional model deviating from the physical laws are fundamentally solved.

[0026] 2. In the remote sensing image disaster dynamic monitoring system of the generated AI, the disaster deduction real-time is based on a neural ODE dynamic deduction engine and a hardware acceleration design, the kinetic equation is optimized for disaster mutation scenes, the minute-level disaster response is realized, the system power consumption is significantly reduced, and the mutation characteristics of the disaster process are effectively captured.

[0027] 3. In the remote sensing image disaster dynamic monitoring system of the generated AI, the economic evaluation dynamic coupling is realized through a building type differentiated conduction coefficient and an economic density factor linkage mechanism, the loss assessment accuracy of a high growth area is significantly improved, the dependence of the traditional scheme on multi-period historical images is broken, and the actuarial process is fully automated.

[0028] 4. The generative AI-based remote sensing image disaster dynamic monitoring system is applied in the disaster insurance claim service scene, star-ground cooperative closed loop control is carried out, a "physical synthesis, dynamic deduction, insurance pricing, temperature drift correction" closed loop system is constructed, temperature drift suppression of star-borne hardware and automatic reinsurance triggering mechanism are carried out, the accuracy of star-borne temperature measurement is greatly improved, the insurance pricing process is efficiently realized, and a monitoring and actuarial integrated solution is formed. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The generative AI-based remote sensing image disaster dynamic monitoring system, device and application overall architecture flowchart disclosed by the embodiments of the present application.

[0030] Figure 2 The hardware acceleration of neural ODE dynamic deduction is compared with CPU deduction.

[0031] Figure 3 The economic density factor dynamic update logic diagram disclosed by the embodiments of the present application. DETAILED DESCRIPTION

[0032] In order to facilitate the understanding of the present application, the present application will be described in more detail below in combination with the drawings and specific embodiments. However, the present application can be realized in many different forms, and is not limited to the embodiments described in the present specification. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosed content of the present application more thorough and comprehensive.

[0033] Embodiment 1: The generative AI-based remote sensing image disaster dynamic monitoring system, device and application overall architecture flowchart, as shown in Figure 1 , includes a data generation layer, an intelligent analysis layer and an application layer.

[0034] I. Data generation layer: by integrating fluid mechanics model and building structure dynamics model, a synthetic data set with physical law label is generated. It includes: physical engine, fluid mechanics simulator, building damage renderer and physical law label binding unit.

[0035] The physical engine adopts a three-dimensional physical engine (such as Unity, Unreal Engine or self-developed engine) to load the digital elevation model (DEM) data of the target area; the DEM data sources can include SRTM (Shuttle Radar Topography Mission), LiDAR surveying and mapping data or satellite stereo image generation data; the engine performs gridding processing according to the terrain slope and surface roughness (Manning coefficient) to provide initial conditions for subsequent fluid simulation.

[0036] A fluid dynamics simulator, inputting terrain DEM data, generates synthetic flood imagery. It adopts the shallow water equation to simulate fluid motion under different terrains, and superimposes rainfall intensity parameters to generate water turbidity characteristics. The fluid simulation adopts a variable terrain Manning coefficient: where n represents the variable terrain Manning coefficient; n_0 represents the initial value; represents the terrain gradient; (x, y) represents the longitude and latitude coordinate values of the terrain gradient.

[0037] A building damage renderer calculates the compression deformation of buildings based on material mechanics parameters, and generates texture details and shadow characteristics of collapsed ruins.

[0038] A physical law label binding unit labels each synthetic data pixel with a fluid velocity field vector and a building structure stress value, and generates a synthetic data set with physical labels.

[0039] II. Intelligent analysis layer: Construct a continuous time domain disaster evolution model, and output the disaster diffusion path and intensity heat map in T hours in the future according to satellite images, unmanned aerial vehicle video streams, weather station data, and synthetic data sets with physical labels. It includes: a three-source spatio-temporal alignment module, a band adaptive attention diffusion model, and a neural differential equation deduction engine.

[0040] The three-source spatio-temporal alignment module receives satellite images, unmanned aerial vehicle video streams, and weather station data, and outputs real small sample data after spatio-temporal alignment of the three.

[0041] The band adaptive attention diffusion model includes a dedicated attention layer for multi-spectral bands (near-infrared band feature enhancement layer, thermal infrared band noise suppression mechanism, and thermal infrared band drift correction unit) and a lightweight band adaptive diffusion architecture. It receives synthetic data and real small sample data, and outputs pixel-level semantic segmentation disaster recognition results.

[0042] The near-infrared band feature enhancement layer amplifies abnormal band features through learnable weights to enhance near-infrared band features: wherein, represents the near-infrared band feature enhancement value; represents the near-infrared band feature; represents the visible light band feature; W_Q represents the query vector weight matrix, and W_K represents the key vector weight matrix; represents the L2 norm operator; d_k represents the key vector dimension, d_k=256 (optimal experimental value); softmax represents the normalization function.

[0043] Thermal infrared band noise suppression mechanism: Utilize weather station temperature data to correct satellite thermal radiation value deviation.

[0044] Thermal infrared band drift correction unit: T_adj = T_raw + 0.15·(T_amb - 25); In the formula, T_adj represents the corrected thermal infrared temperature; T_raw represents the sensor raw output value; T_amb represents the FPGA chip environment temperature.

[0045] Lightweight band adaptive diffusion architecture: the parameter quantity of feature fusion in latent space is less than 30% of that of traditional models.

[0046] Neural differential equation derivation engine: based on satellite, unmanned aerial vehicle, and meteorological station three-source asynchronous input data, a continuous time domain disaster evolution model is constructed, and a future T hour disaster diffusion path and intensity heat map is output;

[0047] A normal differential equation is constructed for disaster diffusion: ; In the formula, h represents a disaster intensity hidden state vector; represents a learnable parameter (including a terrain permeability factor); t represents disaster evolution time.

[0048] A mutation event sensing unit is introduced: when the satellite monitors that the instantaneous rainfall is greater than a threshold, a second-order ODE solidification is automatically switched: ; In the formula, I_rain represents real-time rainfall intensity; γ represents disaster response intensity caused by unit rainfall acceleration; λ represents system inertia (such as delay effect after soil water saturation); h represents a disaster intensity hidden state vector; t represents disaster evolution time.

[0049] The neural differential equation derivation engine is also provided with a derivation result visualization component to generate a three-dimensional diffusion animation with a confidence interval, and supports backtracking of any time node state.

[0050] III. Application layer, including a disaster loss assessment module, the disaster loss assessment module and the distributed storage node and the insurance claim interface establish a signal connection relationship, which receives the derivation result of the intelligent analysis layer, and supports further docking the insurance claim application scene.

[0051] The disaster loss assessment module automatically generates a disaster loss interval valuation report in combination with OpenStreetMap building semantic information and regional economic density atlas, including: a building type classifier, an asset value update engine, and an output layer.

[0052] (1) Building type classifier: receiving GDP growth rate ΔGDP and OpenStreetMap building vector, identifying residential buildings, factories, and farmland based on DeepLabv3+;

[0053] (2) Asset value update engine: access to statistical bureau industry-specific GDP data, distribute loss weights according to building types, and dynamically update economic density factor ρk: [1 + akAGDP] ; where, denotes dynamic economic density factor; denotes baseline economic density factor; AGDP denotes relative change rate of GDP; k denotes building type index; t denotes evolution time; ak is conduction coefficient, which is different for different building types, 0.25 for factory, 0.20 for residence, and 0.18 for commercial building.

[0054] (3) Output layer: output dynamic disaster loss heat map, when pk reaches a certain value, trigger insurance actuarial signal, output claim settlement proposal and automatically trigger reinsurance agreement terms.

[0055] Configure corresponding hardware devices, disaster monitoring edge computing devices:

[0056] (1) Spaceborne FPGA chip: deploy lightweight waveband adaptive attention diffusion model, support on-orbit real-time processing;

[0057] (2) UAV nest linkage interface: when the reasoning engine predicts that the disaster diffusion reaches the threshold area, automatically wake up the UAV cluster;

[0058] (3) Distributed storage unit: store the key frames of disaster evolution and loss assessment results on the chain.

[0059] Disaster insurance service method: receive the geographic fence information of the user's insured area; when the system predicts that the future 24-hour disaster probability of the area is > P%, automatically increase the premium rate, premium floating formula: ; where, denotes dynamic premium (insurance cost adjusted according to real-time disaster risk); denotes basic premium rate; denotes super-threshold disaster probability; h denotes disaster intensity, such as water depth, wind grade, and other key indicators; denotes safety threshold, such as the critical value of disaster intensity that the building can withstand.

[0060] Push the loss assessment report to the insurance company platform within T hours after the disaster occurs.

[0061] The above scheme forms a "physical synthesis-dynamic deduction-insurance pricing-warm drift correction" closed loop scheme, which greatly improves the accuracy of on-board temperature measurement and realizes the efficiency of insurance pricing process through on-board hardware temperature drift suppression and automatic reinsurance triggering mechanism, forming a monitoring and actuarial integrated solution. Among them, the physical synthesis takes the shallow water equation and the material yield strength as the mandatory constraint conditions of the generative AI, so that the synthesized disaster data strictly follows the laws of fluid mechanics and structural mechanics, significantly improves the accuracy of rainstorm prediction and greatly reduces the misjudgment rate of building collapse, and fundamentally solves the technical defects of traditional models deviating from physical laws. Real-time disaster deduction is based on neural differential equation dynamic deduction engine and hardware acceleration design, which optimizes the dynamic equation for disaster mutation scenarios, realizes minute-level disaster response and significantly reduces system power consumption, effectively capturing the mutation characteristics of disaster process. Economic evaluation dynamic coupling through building type differentiated conduction coefficient and economic density factor linkage mechanism, significantly improves the loss assessment accuracy of high growth areas, breaks through the dependence of traditional schemes on multi-period historical images, and realizes the full automation of the actuarial process.

[0062] Embodiment 2: The present application also provides a comparison of hardware acceleration of neural ODE dynamic deduction and CPU deduction. The specific implementation steps of the neural differential equation deduction engine in Embodiment 1: Utilize the second-order ODE solidification technology to achieve efficient deduction through parallel pipeline. In the system scheme, the accurate simulation of the dynamic changes of disasters is realized. In this embodiment, the flood dynamic deduction adopts the modeling method based on neural differential equation, combined with the hardware acceleration processing scheme proposed in the present application, compared with the traditional CPU deduction method, higher timeliness and lower error rate are realized, as shown in Figure 2 .

[0063] In the prior art, the flood deduction of Neural ODE mainly runs on general-purpose CPU architecture, and its process includes:

[0064] S201: The input layer receives the initial state variables (such as water level, rainfall, slope distribution);

[0065] S202: Numerical discretization processing is performed to discretize the ODE continuous model into an iterative form with a time step Δt=10s;

[0066] S203: Explicit Euler or Runge-Kutta method is used for step-by-step iteration;

[0067] S204: It takes about 45 minutes to complete a complete rainstorm event (such as lasting for 1 hour);

[0068] S205: Due to the large time step and limited solution accuracy, the output error is usually more than 30%.

[0069] The application restructures and parallelizes the Neural ODE deduction process by introducing an FPGA / ACAP heterogeneous computing architecture, and realizes the following improvements:

[0070] S206: Input layer initialization

[0071] The following initial variables are received: water depth grid initial value h0(x, y), rainfall rate r(t), terrain slope , building distribution vector diagram.

[0072] S207: Second-order ODE equation solidification

[0073] Based on satellite, unmanned aerial vehicle and weather station three-source asynchronous input data, a Neural ODE structure is introduced to construct a continuous time domain disaster evolution model, and a future T-hour disaster diffusion path and intensity heat map is output. A differential equation for disaster diffusion is constructed: ; In the formula, h represents a disaster intensity hidden state vector; represents a learnable parameter (including a terrain permeability factor); t represents disaster evolution time.

[0074] A mutation event sensing unit is introduced: when the satellite monitors that the instantaneous rainfall is greater than the threshold, the high-order differential equation is automatically switched: ; In the formula, I_rain represents real-time rainfall intensity; γ represents the disaster response intensity caused by unit rainfall acceleration; λ represents system inertia (such as the delay effect after soil water saturation); h represents a disaster intensity hidden state vector; t represents disaster evolution time.

[0075] The spatial domain is divided into multiple sub-blocks (such as 64x64 grids), and ODE calculation is performed in parallel through a multi-channel pipeline. Each channel independently calculates the water depth evolution of its sub-grid, and the state is exchanged between sub-channels through a boundary buffer. A delay-tolerant pipeline scheduling (LTPS) is used to significantly improve hardware utilization and throughput.

[0076] S209: Quickly complete flood evolution deduction

[0077] Under the above optimization, the single rainstorm evolution deduction time is shortened to 10 minutes. Compared with the traditional CPU scheme, the overall deduction time is shortened by about 77.8%.

[0078] S210: Output and error control

[0079] Finally, the flood coverage map, water depth change curve with time and building flooding probability map are output. Compared with the real observation data, the output error of the present scheme is controlled within 9%, which is obviously better than the traditional scheme (>30%).

[0080] Any technical features in the above-described embodiments can be combined in any manner, and, for the sake of brevity, the foregoing description has not described all possible combinations of the technical features. However, it is contemplated that the scope of the disclosure encompasses all possible combinations of the technical features.

Claims

1. A generative AI-based remote sensing image disaster dynamic monitoring system, characterized in that, include: The data generation layer generates a synthetic dataset with physical labels by integrating a fluid dynamics model and a building structure dynamics model. The physical labels include fluid velocity field vectors and building structure stress values. The intelligent analysis layer uses a three-source spatiotemporal alignment module to align satellite imagery, UAV video streams, and weather station data. After the asynchronous input data from these three sources is synchronized, a band-adaptive attention diffusion model outputs semantic segmentation disaster identification results. The band-adaptive attention diffusion model includes a near-infrared band feature enhancement layer with learnable weight amplification, a thermal infrared band noise suppression mechanism based on weather station data, and a correction unit that uses the ambient temperature of the FPGA chip to correct thermal infrared band drift. Then, a continuous-time domain disaster evolution model is constructed through a neural differential equation inference engine, outputting a heat map of the disaster diffusion path and intensity for the next T hours. The neural differential equation inference engine includes a sudden event perception unit that automatically switches to higher-order differential equations when the instantaneous rainfall detected by the satellite exceeds a preset threshold. Thermal infrared band drift correction unit: In the formula, T_adj represents the corrected thermal infrared temperature; T_raw represents the original output value of the sensor; and T_amb represents the ambient temperature of the FPGA chip. At the application layer, combining OpenStreetMap building semantic information with regional economic density maps, a disaster damage range valuation report is automatically generated, and insurance actuarial signals and reinsurance agreement terms are triggered when preset conditions are met.

2. The generative AI-based remote sensing image disaster dynamic monitoring system according to claim 1, characterized in that, The fluid dynamics model uses shallow water equations to simulate fluid motion under different terrains, superimposes rainfall intensity parameters to generate water turbidity characteristics, inputs terrain DEM data, and generates synthetic flood images. During the fluid simulation, a method for calculating the Manning coefficient under varying terrain was employed. In the formula, n represents the Manning coefficient for terrain variation; Indicates the starting value; Indicates terrain gradient; The latitude and longitude coordinates representing the terrain gradient; The building structure dynamics model calculates the compressive deformation of the building based on material mechanics parameters, generating texture details and shadow features of the collapsed ruins.

3. The generative AI-based remote sensing image disaster dynamic monitoring system according to claim 1, characterized in that, Band-adaptive attention diffusion models include: Near-infrared band feature enhancement layer: This layer enhances near-infrared band features by amplifying anomalous features through learnable weights. In the formula, This indicates the enhanced value of the near-infrared band features; Indicates near-infrared band characteristics; Represents visible light band characteristics; W_Q represents the query vector weight matrix, and W_K represents the key vector weight matrix; The L2 norm operator is represented; d_k represents the dimension of the key vector; softmax represents the normalization function. Lightweight band-adaptive attention diffusion architecture: The number of parameters for feature fusion in the latent space is less than 30% of that of traditional models.

4. The generative AI-based remote sensing image disaster dynamic monitoring system according to claim 1, characterized in that, The neural differential equation derivation engine is implemented as follows: Construct the ordinary differential equation for disaster diffusion: In the formula, h represents the hidden state vector of disaster intensity; Indicates the learnable parameter; t represents the disaster evolution time; Introducing a sudden event sensing unit: When the satellite detects that the instantaneous rainfall exceeds a threshold, it automatically switches to higher-order differential equations. In the formula, γ represents the real-time rainfall intensity; γ represents the disaster response intensity caused by a unit rainfall acceleration; λ represents the system inertia; h represents the hidden state vector of disaster intensity; t represents the disaster evolution time. The simulation results visualization component generates a 3D diffusion animation with confidence intervals and supports backtracking to the state at any time point.

5. The generative AI-based remote sensing image disaster dynamic monitoring system according to claim 1, characterized in that, The disaster loss assessment module in the application layer includes: Building type classifier: Receives GDP growth rate ΔGDP and OpenStreetMap building vectors, and identifies residential / factory / farmland based on DeepLabv3+; Asset value update engine: Accesses GDP data by industry from the National Bureau of Statistics, assigns loss weights by building type, and dynamically updates the economic density factor ρk. In the formula, Indicates the dynamic economic density factor; Indicates the benchmark economic density factor; This represents the relative rate of change in GDP; k represents the building type index; t represents the evolution time; αk is the transmission coefficient, which varies for different building types: 0.25 for factories, 0.20 for residential buildings, and 0.18 for commercial buildings. Output layer: Outputs a dynamic disaster loss heat map. When ρk reaches a certain value, it triggers an actuarial signal, outputs a claims recommendation, and automatically triggers reinsurance agreement terms.

6. A disaster monitoring edge computing device, characterized in that, The device integrates a remote sensing image disaster dynamic monitoring system based on generative AI as described in any one of claims 1-5, including: Onboard FPGA chip: Deploys a lightweight band-adaptive attention diffusion model to support real-time on-orbit processing; Drone swarm linkage interface: When the simulation engine predicts that a disaster has spread to the threshold area, the drone swarm is automatically awakened; Distributed evidence storage unit: Stores disaster evolution keyframes and loss assessment results on the blockchain.

7. The application of the remote sensing image disaster dynamic monitoring system based on generative AI as described in any one of claims 1-5, characterized in that, It is applied in disaster insurance claims service scenarios.

8. The application according to claim 7, characterized in that, A closed-loop scheme of "physical synthesis - dynamic deduction - insurance pricing - temperature drift correction" has been formed; Insurance pricing methods and steps: Receive geofence information for the user's insured area; When the system predicts that the probability of disaster in the area in the next 24 hours is greater than P%, the premium rate will be automatically increased. The premium adjustment formula is as follows: In the formula, Indicates dynamic premium; Indicates the basic premium rate; The value represents the probability of a disaster exceeding the threshold; h represents the disaster intensity. Indicates the safety threshold; A damage assessment report will be sent out within T hours after the disaster occurs.

Images