Island-shaped frozen soil degradation prediction method based on remote sensing and ground temperature coupling

Abstract

The invention discloses an island-shaped frozen soil degradation prediction method based on remote sensing and ground temperature coupling, and relates to the technical field of frozen soil degradation prediction in cold regions, and the method comprises the following steps: S1, multi-source data fusion and intelligent decomposition, S2, cross-scale dynamic parameterized model construction, S3, quantum enhanced Bayesian multi-scale calibration, and S4, three-dimensional dynamic uncertainty quantification and prediction output. According to the method, the problems of frozen soil data scarcity and remote sensing conversion precision are solved through multi-source remote sensing and actually measured data fusion in combination with CGAN, AM-PFN and virtual sample closed-loop expansion; the defects of a traditional model are overcome through a microscopic-macroscopic coupling model, DRL parameter switching and meta-learning algorithm selection; through quantum enhanced Bayes, DMD and a four-dimensional uncertainty matrix, cross-scale prediction divergence is reduced; and through a carbon release risk correlation model and a co-evolution product, a prediction dimension is expanded, and a support is provided for ecological protection of the cold region.

Description

Technical Field

[0001] This invention relates to the field of permafrost degradation prediction technology in cold regions, specifically a method for predicting island permafrost degradation based on remote sensing and geothermal coupling. Background Technology

[0002] The core challenge in predicting permafrost degradation lies in quantifying uncertainty. Existing models often focus only on the uncertainty at the level of data noise, neglecting the inherent uncertainty of the model structure itself. Island permafrost, as dispersed permafrost blocks located on the periphery of continuous permafrost zones, in mid-to-low latitude plateaus and high-altitude areas, faces even more complex uncertainty issues in degradation prediction due to its unique characteristics of strong spatial heterogeneity and discontinuous distribution. The structural defects of existing prediction models are the root cause of this predicament. On the one hand, different numerical algorithms (such as the finite element method and the boundary element method) differ fundamentally in their discretization methods, boundary condition handling, and error propagation mechanisms, directly causing systematic biases in the prediction results. On the other hand, models simplify multiple physical processes such as heat conduction, phase change, and hydrological coupling in permafrost through parameterization schemes (e.g., linear or nonlinear assumptions about the relationship between ice and permeability). Such simplifications further amplify biases under climate warming scenarios, failing to reflect the true degradation trend of island permafrost.

[0003] The uncertainties caused by the aforementioned structural defects exhibit a hierarchical nesting characteristic across multiple scales, and practical technical bottlenecks further exacerbate the difficulty of quantification. At the local scale, the sensitivity of numerical algorithms to grid processing and discrepancies in parameterization schemes lead to insufficient stability in the prediction of local island permafrost. At the regional scale, differences in model structure and the coupling with climate forcing factors amplify prediction discrepancies. At the global scale, existing models fail to adequately characterize the spatial non-uniformity and temporal lag effects of permafrost degradation, making it difficult to capture the unique degradation patterns of the dispersed distribution of island permafrost. Furthermore, different models employ heterogeneous data structures and parameterization schemes, lacking interoperability and making it impossible to compare uncertainties across models. The nonlinear coupling of heat conduction and water migration during permafrost thawing leads to a cascading amplification of errors, and the spatiotemporal scales of observed data and model outputs do not match. In addition, the spatial coverage of measured data in permafrost regions is limited, and key data such as ground temperature and soil ice content in island permafrost distribution areas are even scarcer, making it difficult to support model validation and calibration.

[0004] While remote sensing technology enables large-scale permafrost monitoring, its integrated application in predicting permafrost degradation in island areas has significant limitations, becoming a bottleneck in overcoming core challenges: existing models largely rely on static or low-resolution soil parameter data, failing to effectively combine high spatiotemporal resolution remote sensing data to capture dynamic changes in island permafrost; and similar challenges to mixed pixel identification, as seen in canal ice condition research, also exist in remote sensing monitoring of dispersed island permafrost, resulting in insufficient accuracy in converting remote sensing information into model parameters. Therefore, this paper proposes a method for predicting permafrost degradation in island areas based on the coupling of remote sensing and geothermal data to overcome these problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the degradation of island permafrost based on remote sensing and geothermal coupling, so as to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides a method for predicting the degradation of island-type permafrost based on remote sensing and geothermal coupling, comprising the following steps:

[0007] S1. Multi-source data fusion and intelligent decomposition: Multi-source remote sensing data and measured data were collected from the study area. The multi-source remote sensing data included Sentinel-1 SAR data, Sentinel-2 multispectral data, and Landsat-8 TIRS thermal infrared data. The measured data included borehole geothermal stratification data, soil physicochemical property data, meteorological observation data, and CT scan data of permafrost profiles. A conditional generative adversarial network (CGAN) model was constructed to generate a high-resolution surface temperature field. The AM-PFN hybrid pixel decomposition network based on the attention mechanism was used to decompose the multispectral data to obtain the pixel-scale permafrost area proportion (POP) and decomposition confidence. A hybrid pixel uncertainty-heterogeneity transformation model was established to generate the spatial heterogeneity index (UH). Virtual permafrost samples were generated using CGAN and the sample pool was dynamically updated after hierarchical screening and error correction.

[0008] S2. Construction of a cross-scale dynamic parameterization model: Construct a micro-macro coupled constitutive model of permafrost, map the pore structure parameters obtained by CT scans to macro thermal conductivity parameters and establish a dynamic correlation equation between porosity and thermal conductivity; design a parameterization adaptive switching mechanism based on deep reinforcement learning (DRL) to dynamically select parameterization schemes; construct a meta-learning driven multi-algorithm intelligent selection module to match and select the optimal numerical algorithm from the meta-knowledge base;

[0009] S3. Quantum-enhanced Bayesian multi-scale calibration: At the local scale, the grid density and time step are dynamically adjusted using the quantum behavior particle swarm optimization (QPSO) algorithm based on the UH value; at the regional scale, the model parameters are calibrated using a Bayesian network weighted by an attention mechanism with high-resolution surface temperature data as the upper boundary condition; a multi-scale uncertainty propagation network is constructed to input the posterior distribution of local scale parameters as prior knowledge into the regional scale model; and the dynamic model divergence index DMD is defined to quantify the spatiotemporal uncertainty of the model.

[0010] S4. Three-dimensional dynamic uncertainty quantification and prediction output: Develop a four-dimensional uncertainty quantification matrix of parameters, structure, data, and scale, and dynamically update the weights of each dimension through transfer learning; establish a model divergence degree-carbon release risk correlation model to predict carbon source / sink conversion thresholds; generate spatiotemporally continuous island permafrost degradation prediction results and permafrost-vegetation-carbon co-evolution prediction products.

[0011] Furthermore, the CGAN model constructed in step S1 uses a U-Net architecture for its generator and adds a permafrost feature attention module at the decoding end. The discriminator introduces a ground temperature constraint loss function to ensure that the difference between the generated temperature and the measured ground temperature is ≤0.5℃. The final output is a surface temperature field with a spatial resolution of 10m×10m. The borehole ground temperature stratification data is recorded at 1 data point per 10cm depth, the time resolution of the meteorological observation data is 1h, and the resolution of the permafrost profile CT scan data is 50μm.

[0012] Further, in step S1, AM-PFN focuses on permafrost feature regions through a self-attention layer, with the backscattering coefficient of Sentinel-1 SAR data as the weight; AM-PFN decomposes Sentinel-2 multispectral data into four end-member components: permafrost, vegetation, bare soil, and water, with the output decomposition confidence score ranging from 0 to 1, and pixels with a confidence score < 0.6 are marked as needing calibration; the calculation formula for the mixed pixel uncertainty-heterogeneity transformation model is as follows:

[0013] UH=σ(POP)×tanh(k×|T s -T g |);

[0014] in,

[0015] UH is the spatial heterogeneity index;

[0016] σ(POP) is the standard deviation of POP within a 3×3 window;

[0017] T s For generating high-resolution surface temperatures;

[0018] T g This refers to borehole geothermal data;

[0019] k is the sensitivity parameter determined through Bayesian optimization.

[0020] Furthermore, the processing of the virtual permafrost sample in step S1 includes:

[0021] S1.5.1. The improved K-means clustering algorithm is used to classify the credibility of virtual samples. The clustering features include POP value, UH index and CT scan porosity. The virtual samples are divided into three levels: high credibility, medium credibility and low credibility. Low credibility samples are removed.

[0022] S1.5.2 Select high-confidence virtual samples (≥70%) and construct a multiple linear regression equation with the measured borehole data within a 5km radius. The formula is as follows:

[0023] ΔT=c0+c1×POP+c2×UH+c3×R;

[0024] in,

[0025] ΔT is the ground temperature difference value.

[0026] c0-c3 are the regression coefficients.

[0027] R represents the surface roughness;

[0028] S1.5.3 Divide the region into 1km×1km spatial grids. When the number of samples in a certain region is less than 20, trigger CGAN again to generate supplementary samples and repeat steps S1.5.1-S1.5.2 to dynamically update the sample pool.

[0029] Furthermore, in step S2, the micro-macro coupled constitutive model of permafrost uses a 3-layer fully connected network to map the pore structure parameters obtained from CT scans to macroscopic thermal conductivity parameters; the dynamic correlation equation between porosity and thermal conductivity is:

[0030]

[0031] in,

[0032] λ is the thermal conductivity.

[0033] λ0 is the thermal conductivity of pure ice.

[0034] S i For ice saturation,

[0035] T represents temperature.

[0036] a(S i () is the ice saturation correction factor.

[0037] b(T) is the temperature correction factor.

[0038] Furthermore, in step S2, the agent state input of the DRL parameterized adaptive switching mechanism includes the POP value, UH exponent, and historical prediction error; the Actor network outputs the parameterized scheme selection probability; the Critic network uses the dual objective function of minimizing prediction error and maximizing parameter stability, with a weight ratio of 6:4, and updates the network parameters through temporal difference learning, as shown in the formula:

[0039]

[0040] in,

[0041] Q(s,a) is the action value function.

[0042] s represents the current state.

[0043] 'a' represents the selected parameterization scheme.

[0044] r represents the immediate reward.

[0045] γ is the discount factor.

[0046] s′ represents the next state;

[0047] The meta-learning-driven multi-algorithm intelligent selection module adopts the k-nearest neighbor algorithm. Using the mean POP, mean UH, and grid density of the current input data as retrieval features, it matches the five most similar historical scenarios from the meta-knowledge base that stores the performance of finite element method (FEM), boundary element method (BEM), and finite volume method (FVM) in various permafrost scenarios and selects the algorithm with the smallest error as the current calculation algorithm.

[0048] Furthermore, the specific method for local scale calibration in step S3 is as follows:

[0049] An irregular triangular grid was constructed using micro-topographic parameters such as slope and aspect obtained from high-resolution remote sensing. Based on the UH value, the grid density and time step were adjusted using the QPSO algorithm. When UH > 1.0, the grid side length was set to 50m and the time step to 1h; when 0.5 < UH ≤ 1.0, the grid side length was set to 100m and the time step to 6h; when UH ≤ 0.5, the grid side length was set to 200m and the time step to 12h. In regional-scale calibration, an attention-weighted Bayesian network was used to assign weights to network nodes for thermal conductivity, latent heat of phase change, and permeability, with a 30% increase in node weights in high UH regions. The multi-scale uncertainty propagation network was sampled 1000 times using the Markov chain Monte Carlo (MCMC) method to solve for the posterior distribution of parameters. The probabilistic correlation model formula is as follows:

[0050] P(M region |D local )∝P(M local |D local )P(M region |D local );

[0051] in,

[0052] M region These are parameters for the regional-scale model.

[0053] D local For local-scale observation data,

[0054] M local These are the parameters for the local scale model.

[0055] Furthermore, the formula for calculating the dynamic model divergence index DMD in step S3 is as follows:

[0056]

[0057] in,

[0058] M i (t) and M j (t) represents the prediction results of different algorithms at time t.

[0059] ω t The time decay factor,

[0060] C(n,2) is the number of combinations.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] 1. Overcome data bottlenecks and improve the accuracy and data utilization rate of key permafrost parameters inversion:

[0063] To address the scarcity of measured data and insufficient accuracy in remote sensing information conversion in permafrost regions, this study utilizes multi-source data fusion and CGAN-AM-PFN collaborative processing. It integrates Sentinel-1 SAR, Sentinel-2 multispectral, and Landsat-8 TIRS thermal infrared data, along with measured data such as borehole geothermal temperature and CT porosity, to construct a CGAN model with a permafrost feature attention module. This model outputs a high-resolution surface temperature field with a spatial resolution of 10m×10m and a difference of ≤0.5℃ from the measured ground temperature, overcoming the limitation of traditional low-resolution temperature data in characterizing the spatial heterogeneity of island-shaped permafrost. Simultaneously, AM-PFN focuses on permafrost feature areas using SAR backscattering coefficients as weights, decomposing multispectral data into four end-member components: permafrost, vegetation, bare soil, and water. It outputs the pixel-scale permafrost area proportion (POP) and decomposition confidence in the 0-1 range, accurately identifying permafrost information in mixed pixels and avoiding the large bias in identifying dispersed island-shaped permafrost caused by traditional remote sensing decomposition.

[0064] By employing a virtual sample closed-loop expansion mechanism, improved K-means clustering is used to classify the virtual samples generated by CGAN. After removing low-confidence samples, high-confidence samples accounting for ≥70% are selected and combined with borehole measured data within 5km to construct a multiple linear regression equation to correct errors. The sample pool is dynamically updated according to a 1km×1km grid, which effectively alleviates the problem of scarce key data for island permafrost, solves the problem of mismatch between the spatiotemporal scale of observation data and model output, and provides uniform and high-quality sample support for subsequent model calibration.

[0065] 2. Overcome structural defects in the model and enhance its adaptability to complex physical processes in permafrost:

[0066] To address the issue that traditional model parameterization schemes are fixed and cannot reflect the micro-macro coupled physical processes of permafrost, a micro-macro coupled permafrost constitutive model is constructed. Through a three-layer fully connected network, pore structure parameters obtained from CT scans are mapped to macroscopic thermal conductivity parameters. Furthermore, a dynamic correlation equation between porosity and thermal conductivity with ice saturation and temperature correction coefficients is established. This breaks through the limitation of traditional empirical formulas that rely on only a single parameter, and can accurately characterize the dynamic changes of permafrost thermal conductivity with pore structure, ice content, and temperature, avoiding prediction biases caused by simplification of physical processes.

[0067] The design incorporates a DRL parameterized adaptive switching mechanism and a meta-learning multi-algorithm intelligent selection module. The former takes POP value, UH index, and historical prediction error as inputs and dynamically selects the parameterization scheme through an Actor-Critic dual-objective network, avoiding parameter mutations caused by fixed threshold switching. The latter employs the k-nearest neighbor algorithm, using POP mean, UH mean, and grid density as retrieval features to match the optimal algorithm from a meta-knowledge base storing the scenario performance of FEM / BEM / FVM algorithms. This eliminates the need to blindly run all algorithms in parallel, reducing computational costs while minimizing systematic biases caused by differences in discretization methods and boundary handling of different numerical algorithms, thus improving model structural flexibility and prediction stability.

[0068] 3. Achieve coherent quantification of multi-scale uncertainties and reduce cross-scale prediction discrepancies:

[0069] To address the issues of nested multi-scale uncertainties and incomplete quantization in existing models, this paper proposes a quantum-enhanced Bayesian multi-scale calibration method. At the local scale, based on the UH value, the QPSO algorithm dynamically adjusts the grid density and time step to automatically refine the grid in highly heterogeneous regions. At the regional scale, using high-resolution surface temperature as the upper boundary condition, an attention-weighted Bayesian network is used to focus on parameter calibration in highly sensitive areas. Simultaneously, a multi-scale uncertainty propagation network is constructed, using the MCMC method to take the posterior distribution of local parameters as the prior input to the regional model, achieving coherent propagation of uncertainties between scales. This addresses the problems of insufficient characterization of spatial non-uniformity in island-shaped permafrost and amplified regional prediction discrepancies.

[0070] We define a Dynamic Model Divergence (DMD) index and introduce a time decay factor to quantify the prediction divergence of different algorithms in the spatiotemporal dimensions, thus overcoming the shortcomings of traditional models that only statically evaluate divergence. Furthermore, we develop a four-dimensional uncertainty quantification matrix of parameters, structure, data, and scale, and dynamically update the weights of each dimension through transfer learning to comprehensively cover the uncertainties at the parameter, structure, data, and scale levels, avoiding the one-sidedness of existing models that only focus on data noise.

[0071] 4. Expand prediction dimensions, link ecological risks, and provide precise decision support:

[0072] To address the issue that traditional models only predict permafrost degradation and lack ecological correlation, a model correlation model based on model divergence and carbon release risk is established. By analyzing the nonlinear relationship between DMD and remote sensing carbon flux data, the model predicts the carbon source / sink conversion threshold caused by permafrost degradation, directly linking model uncertainty with ecological risk. Simultaneously, a permafrost-vegetation-carbon co-evolution prediction product is generated. Through spatiotemporally coupled heat maps, the nonlinear feedback mechanism of permafrost degradation → vegetation cover change → carbon release is revealed, identifying key ecological regulation nodes and providing precise targets for ecological protection in cold regions, rather than simply outputting permafrost degradation data, thus enhancing the practical application value of the prediction results. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the principle of a method for predicting the degradation of island permafrost based on the coupling of remote sensing and geothermal temperature according to the present invention. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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 are within the scope of protection of the present invention.

[0075] Please see Figure 1 The present invention provides a technical solution:

[0076] This invention aims to address the structural uncertainties, multi-scale error propagation, and spatiotemporal misalignment of data and models in existing island permafrost degradation prediction models. It provides an island permafrost degradation prediction method based on remote sensing and geothermal coupling. Through a defect transformation-dynamic coupling-hierarchical calibration framework, it transforms traditional technical bottlenecks into predictive advantages, achieving high-precision prediction and uncertainty quantification of island permafrost degradation.

[0077] See Figure 1 As shown, an embodiment of a method for predicting island permafrost degradation based on remote sensing and geothermal coupling is presented:

[0078] S1. Multi-source data fusion and intelligent decomposition:

[0079] S1.1 Collect multi-source remote sensing data of the study area, including Sentinel-1 SAR data, Sentinel-2 multispectral data and Landsat-8 TIRS thermal infrared data; at the same time, collect borehole ground temperature measurement data (layered by depth, one data point recorded every 10 cm), soil physicochemical property data (including particle size distribution and organic matter content), meteorological observation data (daily average temperature, precipitation, wind speed, time resolution 1 h) and a small amount of frozen soil profile CT scan data (resolution 50 μm, to obtain pore distribution images).

[0080] S1.2 Construct a Conditional Generative Adversarial Network (CGAN) model. Low-resolution thermal infrared data (30m×30m) and high-resolution multispectral data (10m×10m) are used as inputs. The generator adopts the U-Net architecture and adds a permafrost feature attention module (focusing on pixels in the low-temperature area) at the decoding end. The discriminator introduces a ground temperature constraint loss function (to make the difference between the generated temperature and the measured ground temperature ≤0.5℃). Finally, a high-resolution (10m×10m) surface temperature field is output, which solves the problem of insufficient spatial resolution of traditional methods.

[0081] S1.3. An attention-based hybrid pixel decomposition network (AM-PFN) is proposed. By using a self-attention layer (with the backscattering coefficient of SAR data as the weight to highlight the scattering difference between frozen soil and non-frozen soil) to focus on the characteristic regions of frozen soil, the Sentinel-2 multispectral data is decomposed into four end-member components: frozen soil, vegetation, bare soil and water. The permafrost area ratio (POP) at the pixel scale is obtained, and the decomposition confidence of each end-member is output (between 0 and 1, pixels with confidence < 0.6 are marked as needing calibration).

[0082] S1.4. Establish a mixed pixel uncertainty-heterogeneity transformation model to convert the decomposition error into a spatial heterogeneity quantification index. The calculation formula is as follows:

[0083] UH=σ(POP)×tanh(k×|T s -T g |);

[0084] in,

[0085] UH is the spatial heterogeneity index (with a value of 0-2, and a larger value indicates stronger heterogeneity);

[0086] σ(POP) is the standard deviation of POP within a 3×3 window;

[0087] T s For generating high-resolution surface temperatures;

[0088] T g This is the borehole geothermal data (geothermal values ​​at the corresponding depth).

[0089] k is a sensitivity parameter determined by Bayesian optimization (optimization objective: to make the correlation coefficient of variation of porosity between UH and CT scan ≥ 0.8);

[0090] S1.5. Using the generator module of CGAN, virtual permafrost samples are generated based on a small amount of measured data to expand the training dataset and alleviate the problem of scarce measured data. Specifically, this includes:

[0091] S1.5.1 Virtual Sample Reliability Grading: An improved K-means clustering algorithm is adopted, using POP value, UH index, and CT scan porosity as clustering features (traditional K-means only uses a single numerical feature, while this step combines the multi-dimensional characteristics of permafrost). The generated virtual samples are divided into three levels: high reliability (Euclidean distance between cluster center and measured sample < 0.3), medium reliability (0.3 ≤ distance ≤ 0.6), and low reliability (distance > 0.6). Low reliability samples are directly removed to avoid introducing invalid errors.

[0092] S1.5.2 Closed-loop error correction: Select high-confidence virtual samples (≥70%) and establish an error correction model with the measured borehole data within a 5km radius. Use the difference between virtual sample ground temperature and measured ground temperature as the dependent variable, and POP, UH, and surface roughness (calculated from SRTM data) as independent variables to construct a multiple linear regression equation:

[0093] ΔT=c0+c1×POP+c2×UH+c3×R;

[0094] in,

[0095] ΔT is the ground temperature difference (°C).

[0096] c0-c3 are the regression coefficients (solved using the least squares method, ensuring the corrected error is ≤0.3℃).

[0097] R is the surface roughness (m); the equation is used to correct the ground temperature and ice content parameters of all high-confidence virtual samples, forming a generation-calibration-feedback closed loop;

[0098] S1.5.3 Dynamic Sample Pool Update: Divide the corrected high-confidence virtual samples into regions according to a spatial grid (1km×1km). If the number of samples in each region is less than 20, re-trigger CGAN to generate supplementary samples and repeat steps S151-S152 to ensure uniform spatiotemporal distribution of the sample pool and solve the problem of mismatch between the spatiotemporal scale of the observation data and the model output.

[0099] S2. Construction of cross-scale dynamic parameterization models:

[0100] S2.1 Construct a micro-macro coupled constitutive model of frozen soil, using pore structure parameters (porosity φ, equivalent pore diameter d) obtained from CT scans. p ) Through a deep learning model (using a 3-layer fully connected network, with input layers φ and d) p The output layer (macroscopic thermal conductivity correction coefficient) is mapped to macroscopic thermal conductivity parameters to establish a dynamic correlation equation between porosity and thermal conductivity:

[0101]

[0102] in,

[0103] λ is the thermal conductivity (W / (m·K)).

[0104] λ0 is the thermal conductivity of pure ice (2.2 W / (m·K)).

[0105] S i Ice saturation (0-1),

[0106] T represents temperature (°C).

[0107] a(S i ) is the ice saturation correction coefficient (trained using samples corrected by S15, S i When a = 1, a = 0.8.

[0108] S i When a = 0, a = 0.2.

[0109] b(T) is a temperature correction coefficient (b = -0.1 when T = -10℃, b = -0.5 when T = 0℃), breaking through the limitation of traditional empirical formulas that rely on only a single parameter;

[0110] S2.2 Design a parameterized adaptive switching mechanism based on deep reinforcement learning (DRL). The POP value, UH exponent, and historical prediction error (mean squared error between the model output and the measured data in the first three iterations) are used as inputs to the agent's state. The Actor network outputs the probability of selecting a parameterized scheme (linear model / nonlinear model / freeze-thaw damage model). The Critic network uses "minimizing prediction error + maximizing parameter stability" as its dual objective function (weight ratio 6:4), updating network parameters through temporal difference learning.

[0111]

[0112] in,

[0113] Q(s,a) is the action value function.

[0114] s represents the current state.

[0115] 'a' represents the selected parameterization scheme.

[0116] r represents the immediate reward (the smaller the error and the lower the parameter switching frequency, the larger r is).

[0117] γ is the discount factor (taken as 0.9).

[0118] s′ represents the next state; this avoids parameter abrupt changes caused by traditional fixed threshold switching and reduces model structure uncertainty.

[0119] S2.3 Construct a meta-learning-driven multi-algorithm intelligent selection module. Store the performance of different algorithms (Finite Element Method, Boundary Element Method, and Finite Volume Method) in various permafrost scenarios through a meta-knowledge base. Use the k-nearest neighbor algorithm (k=5) to match the five most similar historical scenarios from the meta-knowledge base with the mean POP, mean UH, and grid density of the current input data as retrieval features. Select the algorithm with the best performance (smallest error) as the current calculation algorithm instead of blindly running all algorithms in parallel, thereby reducing computational costs and minimizing systematic biases in numerical algorithms.

[0120] S3, Quantum-enhanced Bayesian multiscale calibration:

[0121] S3.1 Local Scale (1km) Calibration: Irregular triangular grids are constructed using micro-topographic parameters (slope, aspect) acquired by high-resolution remote sensing. Based on the UH value, the grid density and time step are dynamically adjusted using the Quantum Behavior Particle Swarm Optimization (QPSO) algorithm—when UH > 1.0, the grid side length is set to 50m and the time step to 1h; when 0.5 < UH ≤ 1.0, the grid side length is set to 100m and the time step to 6h; when UH ≤ 0.5, the grid side length is set to 200m and the time step to 12h. Through the probabilistic optimization characteristics of quantum particles, the grid in highly heterogeneous areas is automatically refined, improving the stability of local predictions.

[0122] S3.2 Regional Scale (10km) Calibration: The high-resolution surface temperature data generated in S1.2 is used as the upper boundary condition. Combined with meteorological data (daily average temperature and precipitation) to drive the model, an attention-weighted Bayesian network is adopted. The sample corrected in S1.5 is used as the observation data, and weights are assigned to network nodes (thermal conductivity coefficient, latent heat of phase change, and permeability coefficient)—the weight of nodes in high UH areas is increased by 30%, focusing on parameter calibration in highly sensitive areas and reducing regional scale prediction discrepancies.

[0123] S3.3 Construct a multi-scale uncertainty transmission network, using the local scale calibration results (posterior parameter distribution) as prior knowledge input into the regional scale model, and establish a probabilistic correlation model between scales:

[0124] P(M region |D local )∝P(M local |Dlocal )P(M region |D local );

[0125] in,

[0126] M region For parameters of the region-scale model (such as the region's average thermal conductivity),

[0127] D local This consists of local-scale observation data (corrected virtual samples + measured data).

[0128] M local For the parameters of the local scale model, the posterior distribution is obtained by sampling 1000 times using the Markov chain Monte Carlo (MCMC) method, so as to realize the coherent transmission of multi-scale uncertainty and solve the problem of insufficient characterization of spatial non-uniformity.

[0129] S3.4 Define the Dynamic Model Divergence (DMD) index:

[0130]

[0131] in,

[0132] M i (t) and M j (t) represents the prediction results of different algorithms at time t (e.g., the thickness of permafrost at time t).

[0133] ω t This is a time decay factor (increasing linearly with time; ω = 0.5 at t = 1, ω = 1.0 at t = T, highlighting the impact of recent forecast discrepancies).

[0134] C(n,2) is the number of combinations (C=3 when n=3), which realizes the quantification of spatiotemporal dynamic model uncertainty and makes up for the deficiency of existing models that only statically evaluate discrepancies.

[0135] S4, Three-Dimensional Dynamic Uncertainty Quantification and Innovation Prediction Output

[0136] S4.1 Develop a four-dimensional uncertainty quantification matrix (parameters, structure, data, scale), and introduce transfer learning (using a domain-adaptive network to transfer the uncertainty assessment model from high-data-density regions to low-data-density regions) to dynamically update the weights of each dimension, achieving adaptive quantification of multi-scale uncertainty:

[0137]

[0138] in,

[0139] w s Scale weights (local scale w) s =0.4, regional scale ws =0.6 (adjusted in real time through transfer learning),

[0140] UC p For parameter uncertainty (calculated from the parameter variance after S3 calibration),

[0141] UC s For structural uncertainty (transformed from DMD value),

[0142] UC d For data uncertainty (calculated from the sample confidence level results, UC of high confidence samples), d (Reduce) and comprehensively quantify uncertainty at all levels;

[0143] S4.2 Establish a model divergence degree-carbon release risk correlation model. By analyzing the nonlinear relationship between DMD and remote sensing-derived carbon flux data (co-derived from Sentinel-2 NDVI and Landsat-8 surface temperature), predict the carbon source / sink conversion threshold caused by permafrost degradation. The formula is as follows:

[0144] CR=g(DMD,NDVI,T)=α×DMD β ×NDVI γ ×e δT ;

[0145] in,

[0146] CR stands for Carbon Release Risk Index (>1 indicates a carbon source, <1 indicates a carbon sink).

[0147] NDVI is the vegetation index (0-1), with α = 0.8, β = 1.2, γ = -0.5, and δ = 0.1.

[0148] To establish the correlation between model structural uncertainty and ecological risk using parameters obtained through training with corrected samples, and to expand the prediction dimensions;

[0149] S4.3 Generate spatiotemporally continuous island permafrost degradation prediction results, including permafrost thickness changes over the next 5 / 10 / 20 years (spatial resolution 10m×10m), active layer depth (temporal resolution 1 year), and degradation probability distribution (classified as high risk >70%, medium risk 30%-70%, and low risk <30%), along with the confidence interval of each prediction result (calculated by UC value).

[0150] S4.4 Output permafrost-vegetation-carbon co-evolution prediction products. By drawing a spatiotemporal coupled heat map of the three (the horizontal axis is longitude, the vertical axis is latitude, and the color represents the intensity of co-evolution), the nonlinear feedback mechanism of permafrost degradation → vegetation cover change → carbon release is revealed. Key ecological regulation nodes of permafrost degradation (such as areas with NDVI = 0.3 and T = -2℃ are carbon release sensitive nodes) are identified, providing precise targets for ecological protection in cold regions.

[0151] Summarize:

[0152] Multi-source data integration + CGAN-AM-PFN collaborative processing: Integrating Sentinel-1 / 2 and Landsat-8 remote sensing data, as well as measured data such as layered borehole geothermal temperature and CT porosity; using CGAN with a permafrost feature attention module to output 10m high-resolution surface temperature; and AM-PFN decomposing mixed pixels using SAR backscattering coefficients as weights, outputting the permafrost area proportion (POP) and confidence level. This addresses the problems of remote sensing relying on static low-resolution data, difficulty in identifying mixed pixels, and insufficient accuracy in converting remote sensing information into model parameters.

[0153] Virtual sample closed-loop expansion: High-confidence virtual samples are selected by hierarchical screening using an improved K-means method (characterized by POP, UH, and CT porosity), and then corrected using multiple linear regression. The sample pool is dynamically updated according to a 1km grid. This addresses the problems of scarce measured data in permafrost regions, insufficient key data for island permafrost, and mismatch between the spatiotemporal scale of observational data and model output.

[0154] A micro-macro coupled constitutive model maps CT pore parameters (porosity, equivalent diameter) to macroscopic thermal conductivity through a three-layer fully connected network, establishing dynamic correlation equations for ice saturation and temperature correction coefficients. This addresses the problems of simplified models of permafrost heat conduction, simplistic parameterization schemes, and inability to reflect real heat conduction patterns.

[0155] DRL Adaptive Parameter Switching + Meta-Learning Algorithm Selection: Using POP, UH, and historical error as inputs, DRL dynamically selects the parameterization scheme through a dual objective (minimum error + parameter stability); based on a meta-knowledge base (storing algorithm scenario performance), it uses k-nearest neighbors to select the optimal numerical algorithm (FEM / BEM / FVM). This addresses the issues of fundamental differences between different numerical algorithms, systematic biases caused by simplification of parameterization schemes, and neglect of model structure uncertainties.

[0156] Multi-scale quantum-enhanced Bayesian calibration: At the local scale, QPSO dynamically adjusts the grid density and time step according to the UH value; at the regional scale, attention-weighted Bayesian networks (with 30% weight increase in high UH regions) are used to calibrate parameters; and MCMC is used to transfer the posterior of local parameters as the prior to the regional scale. This addresses the problems of nested multi-scale uncertainties, insufficient local prediction stability, amplified regional prediction discrepancies, and inadequate characterization of spatial non-uniformity.

[0157] Dynamic Model Divergence (DMD) metric: Introduces a time decay factor (highlighting recent divergence) to quantify the predictive divergence of different algorithms in the spatiotemporal dimensions. This addresses the lack of dynamic quantification of model structural uncertainty and the problem that existing models only statically evaluate divergence.

[0158] A four-dimensional uncertainty quantification matrix is ​​constructed by creating a parameter-structure-data-scale matrix and dynamically updating the weights through transfer learning (transferring models from high-density areas to low-density areas) to comprehensively quantify uncertainty across all dimensions. This addresses the problem of existing models focusing only on data noise-level uncertainty and providing incomplete uncertainty quantification.

[0159] The DMD-Carbon Release Risk Association Model establishes a nonlinear relationship between DMD and remote sensing-derived carbon flux, predicting carbon source / sink conversion thresholds (CR>1 for carbon sources, <1 for carbon sinks). This addresses the issues of single prediction dimensions and lack of correlation with the ecological impacts of permafrost degradation, expanding the application value of prediction.

Claims

1. A method for predicting the degradation of island-type permafrost based on remote sensing and geothermal coupling, characterized in that, Includes the following steps: S1. Multi-source data fusion and intelligent decomposition: Multi-source remote sensing data and measured data were collected from the study area. The multi-source remote sensing data included Sentinel-1 SAR data, Sentinel-2 multispectral data, and Landsat-8 TIRS thermal infrared data. The measured data included borehole geothermal stratification data, soil physicochemical property data, meteorological observation data, and CT scan data of permafrost profiles. A conditional generative adversarial network (CGAN) model was constructed to generate a high-resolution surface temperature field. The AM-PFN hybrid pixel decomposition network based on the attention mechanism was used to decompose the multispectral data to obtain the pixel-scale permafrost area proportion (POP) and decomposition confidence. A hybrid pixel uncertainty-heterogeneity transformation model was established to generate the spatial heterogeneity index (UH). Virtual permafrost samples were generated using CGAN and the sample pool was dynamically updated after hierarchical screening and error correction. S2. Construction of a cross-scale dynamic parameterization model: Construct a micro-macro coupled constitutive model of permafrost, map the pore structure parameters obtained by CT scans to macro thermal conductivity parameters and establish a dynamic correlation equation between porosity and thermal conductivity; design a parameterization adaptive switching mechanism based on deep reinforcement learning (DRL) to dynamically select parameterization schemes; construct a meta-learning driven multi-algorithm intelligent selection module to match and select the optimal numerical algorithm from the meta-knowledge base; S3. Quantum-enhanced Bayesian multi-scale calibration: At the local scale, the grid density and time step are dynamically adjusted using the quantum behavior particle swarm optimization (QPSO) algorithm based on the UH value; at the regional scale, the model parameters are calibrated using a Bayesian network weighted by an attention mechanism with high-resolution surface temperature data as the upper boundary condition; a multi-scale uncertainty propagation network is constructed to input the posterior distribution of local scale parameters as prior knowledge into the regional scale model; and the dynamic model divergence index DMD is defined to quantify the spatiotemporal uncertainty of the model. S4. Three-dimensional dynamic uncertainty quantification and prediction output: Develop a four-dimensional uncertainty quantification matrix of parameters, structure, data, and scale, and dynamically update the weights of each dimension through transfer learning; establish a model divergence degree-carbon release risk correlation model to predict carbon source / sink conversion thresholds; generate spatiotemporally continuous island permafrost degradation prediction results and permafrost-vegetation-carbon co-evolution prediction products.

2. The island permafrost degradation prediction method based on remote sensing and geothermal coupling as described in claim 1, characterized in that: The CGAN model constructed in step S1 uses a U-Net architecture for its generator and adds a permafrost feature attention module to the decoder. The discriminator introduces a ground temperature constraint loss function to ensure that the difference between the generated temperature and the measured ground temperature is ≤0.5℃. The final output is a surface temperature field with a spatial resolution of 10m×10m. The borehole ground temperature stratification data is recorded at 1 data point every 10cm depth, the time resolution of the meteorological observation data is 1h, and the resolution of the permafrost profile CT scan data is 50μm.

3. The island permafrost degradation prediction method based on remote sensing and geothermal coupling as described in claim 1, characterized in that: In step S1, AM-PFN focuses on the permafrost feature area through a self-attention layer, with the backscattering coefficient of Sentinel-1 SAR data as the weight. AM-PFN decomposes Sentinel-2 multispectral data into four end-member components: permafrost, vegetation, bare soil, and water. The output decomposition confidence level ranges from 0 to 1, and pixels with a confidence level < 0.6 are marked as needing calibration. The calculation formula for the mixed pixel uncertainty-heterogeneity transformation model is as follows: UH=σ(POP)×tanh(k×|T s -T g |); in, UH is the spatial heterogeneity index; σ(POP) is the standard deviation of POP within a 3×3 window; T s For generating high-resolution surface temperatures; T g This refers to borehole geothermal data; k is the sensitivity parameter determined through Bayesian optimization.

4. The island permafrost degradation prediction method based on remote sensing and geothermal coupling as described in claim 1, characterized in that: The processing of virtual permafrost samples in step S1 includes: S1.5.

1. The improved K-means clustering algorithm is used to classify the credibility of virtual samples. The clustering features include POP value, UH index and CT scan porosity. The virtual samples are divided into three levels: high credibility, medium credibility and low credibility. Low credibility samples are removed. S1.5.2 Select high-confidence virtual samples (≥70%) and construct a multiple linear regression equation with the measured borehole data within a 5km radius. The formula is as follows: ΔT=c0+c1×POP+c2×UH+c3×R; in, ΔT is the ground temperature difference value. c0-c3 are the regression coefficients. R represents the surface roughness; S1.5.3 Divide the region into 1km×1km spatial grids. When the number of samples in a certain region is less than 20, trigger CGAN again to generate supplementary samples and repeat steps S1.5.1-S1.5.2 to dynamically update the sample pool.

5. The island permafrost degradation prediction method based on remote sensing and geothermal coupling as described in claim 1, characterized in that: In step S2, the micro-macro coupled constitutive model of permafrost uses a 3-layer fully connected network to map the pore structure parameters obtained from CT scans to macroscopic thermal conductivity parameters; the dynamic correlation equation between porosity and thermal conductivity is: in, λ is the thermal conductivity. λ0 is the thermal conductivity of pure ice. S i For ice saturation, T represents temperature. a(S i () is the ice saturation correction factor. b(T) is the temperature correction factor.

6. The island permafrost degradation prediction method based on remote sensing and geothermal coupling as described in claim 1, characterized in that: In step S2, the agent state inputs of the DRL parameterized adaptive switching mechanism include the POP value, UH exponent, and historical prediction error; the Actor network outputs the parameterized scheme selection probability; the Critic network uses the dual objective function of minimizing prediction error and maximizing parameter stability, with a weight ratio of 6:4, and updates the network parameters through temporal difference learning, as shown in the formula: in, Q(s,a) is the action value function. s represents the current state. 'a' represents the selected parameterization scheme. r represents the immediate reward. γ is the discount factor. s′ represents the next state; The meta-learning-driven multi-algorithm intelligent selection module adopts the k-nearest neighbor algorithm. Using the mean POP, mean UH, and grid density of the current input data as retrieval features, it matches the five most similar historical scenarios from the meta-knowledge base that stores the performance of finite element method (FEM), boundary element method (BEM), and finite volume method (FVM) in various permafrost scenarios and selects the algorithm with the smallest error as the current calculation algorithm.

7. The island permafrost degradation prediction method based on remote sensing and geothermal coupling as described in claim 1, characterized in that: The specific method for local scale calibration in step S3 is as follows: An irregular triangular grid was constructed using micro-topographic parameters such as slope and aspect obtained from high-resolution remote sensing. Based on the UH value, the grid density and time step were adjusted using the QPSO algorithm. When UH > 1.0, the grid side length was set to 50m and the time step to 1h; when 0.5 < UH ≤ 1.0, the grid side length was set to 100m and the time step to 6h; when UH ≤ 0.5, the grid side length was set to 200m and the time step to 12h. In regional-scale calibration, an attention-weighted Bayesian network was used to assign weights to network nodes for thermal conductivity, latent heat of phase change, and permeability, with a 30% increase in node weights in high UH regions. The multi-scale uncertainty propagation network was sampled 1000 times using the Markov chain Monte Carlo (MCMC) method to solve for the posterior distribution of parameters. The probabilistic correlation model formula is as follows: P(M region |D local )∝P(M local |D local )P(M region |D local ); in, M region These are parameters for the regional-scale model. D local For local-scale observation data, M local These are the parameters for the local scale model.

8. The island permafrost degradation prediction method based on remote sensing and geothermal coupling as described in claim 1, characterized in that: The formula for calculating the dynamic model divergence index DMD in step S3 is as follows: in, M i (t) and M j (t) represents the prediction results of different algorithms at time t. ω t The time decay factor, C(n,2) is the number of combinations.

Images