Forest canopy and surface fuel moisture content inversion method and device based on sar remote sensing and medium

Abstract

The application discloses a forest canopy and surface combustible moisture content inversion method based on SAR remote sensing, which comprises the following steps: 1) polarization decomposition, using preprocessed Sentinel-1 IW SLC data, extracting canopy characteristic parameters and surface characteristic parameters through Stokes decomposition; 2) using a water cloud model to preliminarily estimate the forest combustible moisture content, first obtaining an optimal parameter set of the water cloud model, and then using a particle swarm optimization algorithm to inversely obtain the forest combustible moisture content; and 3) error correction, constructing a machine learning model to non-linearly correct the error in the initial forest combustible moisture content estimation result obtained in step 2). The forest canopy and surface combustible moisture content inversion method based on SAR remote sensing realizes synchronous estimation of LFMC and DFMC, and effectively solves the problem of vertical layering estimation of FMC for a long time.

Description

Technical Field

[0001] This invention relates to the field of forest fire prevention, and in particular to a method, apparatus and medium for the simultaneous inversion of forest canopy active combustible moisture content (LFMC) and surface dead combustible moisture content (DFMC). Background Technology

[0002] In recent years, global climate change has led to frequent forest fires around the world, causing severe ecological damage and carbon emissions. As a core parameter for wildfire risk assessment, forest fuel moisture content (FMC) directly affects the flammability of vegetation and the rate of fire spread. FMC is defined as the ratio of the difference between the fresh weight and dry weight of a vegetation sample to the dry weight. It is usually expressed as a percentage.

[0003] Due to the significant spatiotemporal heterogeneity of forest fire content (FMC), its monitoring requires high-frequency and efficient methods. Currently, large-scale FMC estimation mainly relies on optical remote sensing imagery. However, the acquisition of optical imagery is susceptible to weather conditions and cloud cover, and its limited penetration hinders the acquisition of crucial information beneath the forest canopy. Studies have shown that forest combustibles are distributed vertically in layers, with understory shrubs and surface litter layers exhibiting higher fire sensitivity. Even small fluctuations in the surface DFMC within these layers can significantly impact overall fire risk. Therefore, to achieve accurate wildfire risk assessment, it is essential to simultaneously estimate both canopy LFMC and surface DFMC and clearly understand their vertical distribution characteristics.

[0004] Most current research focuses on estimating a single parameter, namely surface DFMC. Simultaneously quantifying canopy LFMC and surface DFMC requires acquiring independent data and employing specific estimation procedures, leading to inefficiency and difficulty in achieving spatiotemporal consistency. Synthetic Aperture Radar (SAR), with its all-weather, all-time observation capabilities, sensitivity to moisture changes, and ability to penetrate vegetation canopies, provides a technical pathway for the simultaneous estimation of LFMC and DFMC. Theoretically, different vegetation structures (such as canopy and surface) exhibit different scattering mechanisms in polarimetric SAR data, providing a physical basis for using polarimetric decomposition techniques to separate canopy and surface scattering signals, thus possessing the potential for simultaneous estimation of LFMC and DFMC. Although some studies have attempted to apply SAR to LFMC estimation in recent years, an effective method for simultaneously estimating DFMC and LFMC still lacks a clear approach.

[0005] Currently, SAR-based FMC estimation methods are mainly divided into two categories: semi-empirical models and machine learning models. Semi-empirical models are based on physical mechanisms and calibrated using experimental data, possessing strong interpretability and theoretical support. They can still provide reliable results even with limited data and exhibit good generalization ability within their assumptions. However, these models rely on specific simplifications and assumptions, limiting their effectiveness in handling complex nonlinear relationships, and their accuracy may be constrained by model approximation errors. In contrast, machine learning models excel at automatically learning from large-scale data, effectively capturing complex, high-dimensional, and nonlinear relationships; given sufficient high-quality data, they can achieve high-precision estimations. However, these models are highly dependent on the quantity and quality of training data, lack explicit physical meaning, and may suffer from insufficient generalization ability when faced with changes in data distribution. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method, device and medium for inverting the moisture content of forest canopy and surface combustibles based on SAR remote sensing.

[0007] To solve the above-mentioned technical problems, the technical solution proposed in this invention is: a method for inverting the moisture content of forest canopy and surface combustibles based on SAR remote sensing, comprising the following steps:

[0008] 1) Polarization decomposition: Using preprocessed Sentinel-1 IW SLC data, canopy feature parameters are extracted through Stokes decomposition. and surface characteristic parameters ;

[0009] 2) Preliminary estimation of forest fuel water content (FMC) using the Water Cloud Model (WCM): First, the optimal parameter set of the Water Cloud Model (WCM) is obtained through a two-stage optimization strategy, and then the forest fuel water content (FMC) is obtained by inversion using the Particle Swarm Optimization (PSO) algorithm.

[0010] 3) Error correction: Using SAR data, radar incident angle, slope and aspect as inputs, a machine learning model is constructed to perform nonlinear correction on the error in the initial forest fuel moisture content (FMC) estimation results obtained in step 2).

[0011] In the aforementioned method for inverting forest canopy and surface combustible moisture content based on SAR remote sensing, preferably, in step 1), canopy characteristic parameters are extracted using Stokes decomposition. and surface characteristic parameters It uses the polarization Stokes matrix to decompose the scattered wave into and Two polarization parameters represent the scattering components of the forest canopy and the dihedral scattering from the ground surface, respectively; their specific relationships are as follows:

[0012] ;

[0013] in and The model parameters for surface scattering can be obtained using Stokes vectors;

[0014] .

[0015] In the aforementioned method for inverting forest canopy and surface combustible water content based on SAR remote sensing, preferably, the Stokes vector of the scattered wave is entirely determined by the C2 matrix, specifically expressed as:

[0016] .

[0017] The above-mentioned method for inverting the moisture content of forest canopy and surface combustibles based on SAR remote sensing, preferably, includes one or more of the following preprocessing steps in step 1): orbit correction, radiometric calibration, pulse band splicing, polarization matrix generation, multi-view, terrain smoothing, polarization spot filtering, and terrain correction.

[0018] The above-mentioned method for inverting forest canopy and surface combustible water content based on SAR remote sensing is preferably obtained by using a two-stage optimization strategy in step 2) to obtain the optimal parameter set of the water cloud model (WCM). The two-stage optimization strategy includes a Monte Carlo pre-search stage and a least squares fine optimization stage.

[0019] The Monte Carlo method pre-search stage involves generating 5000 sets of random number combinations in the interval (0, 10) for both the canopy and the ground surface. Each set of the canopy contains 4 numbers, and each set of the ground surface contains 6 numbers. These random parameter combinations are used as random parameter combinations to optimize the parameters. The objective function is solved using these 5000 sets of random parameter combinations, and the optimal parameter combination is recorded.

[0020] The least squares fine optimization stage involves selecting the optimal parameter combination from the pre-search stage, solving the nonlinear least squares problem through a trust region adjustment strategy, and obtaining the optimal model parameters.

[0021] The above-mentioned method for inverting forest canopy and surface combustible moisture content based on SAR remote sensing is preferably wherein the forest combustible moisture content in step 2) includes forest canopy combustible moisture content and forest surface combustible moisture content.

[0022] ;

[0023] ;

[0024] in, and LFMC and DFMC represent the optimal canopy and surface combustible water content, respectively, and N represents the number of observation samples.

[0025] In the above-mentioned method for inverting forest canopy and surface combustible moisture content based on SAR remote sensing, preferably, the input matrix of the machine learning model in step 3) is:

[0026] ;

[0027] Where VV and VH represent backscattering with different polarization modes. The radar incident angle;

[0028] The model output target y is the difference between the FMC derived from the semi-empirical model inversion and the measured FMC.

[0029] .

[0030] A device for retrieving forest canopy and surface combustible moisture content based on SAR remote sensing includes a processor, which executes the above-mentioned method when running a program.

[0031] A computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the above-described method for inverting the moisture content of forest canopy and surface combustibles based on SAR remote sensing.

[0032] Compared with the prior art, the advantages of the present invention are as follows: the forest canopy and surface combustible water content inversion method based on SAR remote sensing of the present invention realizes the simultaneous estimation of LFMC and DFMC, effectively solving the long-standing problem of vertical stratification estimation of FMC.

[0033] The SAR remote sensing-based forest canopy and surface combustible water content inversion method of this invention achieves three key breakthroughs: (1) Dual-parameter synchronous estimation capability: By effectively separating the canopy and surface scattering components through Stokes decomposition, the method in this embodiment only requires a single SAR image to synchronously invert LFMC and DFMC. This breakthrough provides a reliable technical path for generating spatiotemporally consistent FMC distribution maps, significantly improving the accuracy and timeliness of dynamic wildfire risk assessment.

[0034] (2) Advantages of innovative hybrid modeling: By combining semi-empirical models with machine learning error correction, the framework retains the constraints of physical mechanisms while introducing nonlinear error compensation capabilities. This hybrid modeling strategy not only maintains the physical interpretability of the model but also significantly improves the accuracy and stability of parameter estimation.

[0035] (3) Good potential for operational application: The method is highly compatible with the existing Sentinel-1 satellite data system, and the framework adopts a modular architecture, which has strong adaptability and can be widely applied to various forest types and geographical environments. Attached Figure Description

[0036] Figure 1 This is a map showing the distribution of the study area and ground measurement stations in Example 1.

[0037] Figure 2 This is a flowchart of the method for inverting the moisture content of forest canopy and surface combustibles based on SAR remote sensing in Example 1.

[0038] Figure 3 This is a scattering map of the Sentinel-1 image polarization decomposition into mv in Example 1.

[0039] Figure 4 This is a scattering map of the Sentinel-1 image polarization decomposed into ms in Example 1.

[0040] Figure 5 This is a heatmap showing the correlation between LFCM and meteorological and remote sensing variables.

[0041] Figure 6 This is a heatmap showing the correlation between DFMC and meteorological and remote sensing variables.

[0042] Figure 7 Comparison of RMSE and time consumption for PSO, GA, and SA optimization algorithms in solving the canopy LFMC for water cloud models.

[0043] Figure 8 Comparison of RMSE and time consumption for PSO, GA, and SA optimization algorithms in solving the canopy DFMC for water cloud models.

[0044] Figure 9 The relationship between the estimated LFMC based on the decomposed Sentinel-1 data and the measured value.

[0045] Figure 10 The relationship between the estimated DFMC based on the decomposed Sentinel-1 data and the measured value.

[0046] Figure 11 The results are LFMC estimations obtained using the method described in Example 1.

[0047] Figure 12 The results are DFMC estimations obtained using the method described in Example 1.

[0048] Figure 13 The estimated monthly average LFMC in California for 2024 using the method of Example 1.

[0049] Figure 14 The estimated monthly average DFMC in California for 2024 using the method of Example 1. Detailed Implementation

[0050] To facilitate understanding of the present invention, the present invention will be described more fully and in detail below with reference to preferred embodiments, but the scope of protection of the present invention is not limited to the following specific embodiments.

[0051] It should be noted that when a component is described as being "fixed to, attached to, connected to or connected to" another component, it can be directly fixed to, attached to, connected to or connected to the other component, or it can be indirectly fixed to, attached to, connected to or connected to the other component through other intermediate connectors.

[0052] Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the scope of the invention.

[0053] Example 1

[0054] The study area and data sources for the SAR remote sensing-based forest canopy and surface combustible moisture content inversion method in Example 1 are as follows:

[0055] 1. Study Area and Data

[0056] 1.1 Study Area

[0057] The study area in this embodiment, such as Figure 1As shown, California (32°30′N–42°00′N, 114°8′W–124°24′W), located on the southwestern Pacific coast of the United States, is the third largest state in the US, with a total area of ​​approximately 424,000 square kilometers. Its diverse terrain encompasses the Coast Mountains, the Central Valley, the Sierra Nevada Mountains, and desert regions. The region has a typical Mediterranean climate, with hot, dry summers and mild, rainy winters. Uneven annual precipitation distribution leads to significant water loss from vegetation during the dry season (typically May to October), resulting in a significant decrease in forest cover (FMC). While the forest cover is 32.71%, which is not particularly high overall, the total forest area still exceeds 138,600 square kilometers. Due to the high overlap between California's dry season and hot season, wildfires are frequent and have shown a marked increase in recent decades. Between 1972 and 2018, the average annual burned area in California increased fivefold, with summer wildfire burned area increasing more than eightfold (WILLIAMS AP, ABATZOGLOU JT, GERSHUNOV A, et al. Observed Impacts of Anthropogenic Climate Change on Wildfire in California [J]. Earth's Future, 2019, 7(8): 892–910.). This study selected 33 FMC ground observation stations in California, covering both active canopy fuels and dead surface fuels, and their distribution is as follows: Figure 1 As shown.

[0058] 1.2 Data Preprocessing

[0059] The data in this embodiment includes measured data from Sentinel-1 IW SLC, SRTM DEM, MODIS LAI, and FMC; the specific sources and data preprocessing are as follows:

[0060] 1.2.1 Sentinel-1 IW SLC

[0061] Sentinel-1 is a synthetic aperture radar (SAR) satellite under the European Space Agency's (ESA) Copernicus program. It consists of Sentinel-1A (launched in 2014) and Sentinel-1B (launched in 2016, now retired), providing all-weather, all-time radar remote sensing data. SLC (Single Look Complex) is a high-precision data product from Sentinel-1, preserving the original amplitude and phase information of the radar signal, suitable for advanced applications such as interferometry (InSAR) and polarization analysis. The data format is complex, containing both real and imaginary parts, supporting phase analysis, and offering two modes: IW (wide-swath interferometric mode) and SM (strip mode). This study selected IW, acquiring the data through the Alaska Satellite Facility Data Search Vertex, and preprocessing it in SNAP to generate a C2 matrix.

[0062] 1.2.2 SRTM DEM

[0063] The Shuttle Radar Topography Mission (SRTM) was an international project conducted by NASA, the National Geospatial-Intelligence Agency (NGA), and the space agencies of Germany and Italy. Launched in February 2000 aboard the Space Shuttle Endeavour, it conducted high-precision topographic mapping of the globe using a radar system. SRTM data covered land areas between 60°N and 56°S latitude, providing the most complete digital elevation model (DEM) at the time. It was available in two spatial resolutions: 1 arcsecond (approximately 30 meters) and 3 arcseconds (approximately 90 meters). This study selected the SRTM Plus product, provided by NASA JPL, with a resolution of 1 arcsecond (approximately 30 meters). SRTM data was acquired from Google Earth Engine (GEE), and slope and aspect were calculated within GEE.

[0064] 1.2.3, MODIS LAI

[0065] The Leaf Area Index (LAI) product used in this study is one of NASA's MODIS products, named MCD15A3H. This dataset is a 4-day synthetic dataset from MODIS Collection 6.1 (version "061"). Generated from joint observation data from Terra and Aqua satellites, it has a spatial resolution of 500 meters, is provided in HDF4 format, uses a sinusoidal iso-area projection, and is georeferenced based on the WGS84 coordinate system. To ensure spatial alignment with Sentinel-1 data, the MCD15A3H product was resampled using GEE, thus achieving consistent integration and analysis between the two datasets. MCD15A3H provides global leaf area index (LAI) estimates every 4 days and is primarily used in areas such as surface process modeling, carbon cycle research, vegetation dynamics monitoring, and ecosystem productivity assessment. LAI represents the ratio of total leaf area to leaf area cover per unit area of ​​land. Since the dataset was generated by selecting the best pixels from all data acquired over four days, and the LAI does not change much in the short term, when the data was not collected on the same day, the closest date in time was selected for time alignment.

[0066] 1.2.4 FMC Measured Data

[0067] The ground-based observational data used in this study came from the National Fuel Moisture Database (NFMD). This database covers 976 fixed monitoring stations, primarily in the western United States. Sampling was conducted in the afternoon on clear, rain-free, and dew-free days to minimize the impact of meteorological conditions on sample moisture content. In practice, after collecting target vegetation samples, researchers used the drying and weighing method to determine the fuel moisture content (FMC), calculating the FMC value by comparing the mass difference before and after drying. Due to limitations imposed by satellite transit conditions, the presence of the Sentinel-1 satellite on the day of ground sampling was used as a selection criterion, resulting in 400 valid observational data points from 33 monitoring stations in California for this analysis. Of these, 200 were canopy LFMC data and 200 were surface DFMC data. In the canopy and surface data, 120 unique integers within the range (0, 200) were randomly generated as index numbers to select data for water cloud model parameter calibration and subsequent machine learning model training. The remaining samples were used as a test set to independently evaluate the model's generalization ability.

[0068] 2. Inversion method for forest canopy and surface combustible moisture content based on SAR remote sensing

[0069] like Figure 2 The method for retrieving forest canopy and surface combustible moisture content based on SAR remote sensing, as shown, includes the following steps:

[0070] 2.1 Polarization decomposition;

[0071] 2.2. Preliminary estimation and steps for forest fuel moisture content (FMC) using the water cloud model (WCM);

[0072] 2.3 Error correction.

[0073] 2.1 Polarization decomposition: Using preprocessed Sentinel-1 IW SLC data, canopy feature parameters were extracted through Stokes decomposition. and surface characteristic parameters As mentioned above, the preprocessing in step 1) includes orbit correction, radiometric calibration, pulse band splicing, polarization matrix generation, multi-view, terrain smoothing, polarization speckle filtering, and terrain correction. These eight preprocessing methods are all known technical solutions in the current technology.

[0074] The following are descriptions of the 8 preprocessing methods:

[0075] (1) Orbit Correction (Apply-Orbit-File): Using the precise orbit file released by the European Space Agency, the geometric positioning accuracy of the image is improved to the centimeter level, laying the foundation for subsequent terrain correction and multi-temporal registration.

[0076] (2) Radiation Calibration: This step converts the raw digital values ​​(DN) into radar backscattered signals with physical meaning. The output format of this step needs to be selected as complex, because complex data is a prerequisite for the subsequent construction of polarization covariance or coherence matrix.

[0077] (3) Deburst: stitches multiple sub-strips into a continuous and seamless complete image, eliminating spatial breaks caused by scanning modes and ensuring the continuity and integrity of subsequent processing.

[0078] (4) Generate the polarization matrix (Polarimetric-Speckle-Filter): For dual polarization, construct the second-order covariance matrix C2. This matrix records the intensity and phase relationship between each polarization channel in complex form and is the core input for advanced polarization analysis such as polarization decomposition, target feature extraction, and land cover classification.

[0079] (5) Multilook: In order to suppress the inherent speckle noise of SAR and balance the resolution and signal-to-noise ratio, the pixels are averaged in the range and azimuth directions to match the incident angle distribution in Sentinel-1 IW mode. Although the spatial resolution of the multilook image is reduced, the smoothness and stability of the image are significantly improved.

[0080] (6) Terrain-Flattening: Using an external digital elevation model to correct the local incident angle of the image, eliminating the distortion of the backscattering coefficient caused by the terrain slope (such as upslope enhancement and downslope attenuation), so that the signal values ​​of different terrain areas are physically comparable.

[0081] (7) Polarization speckle filter: An adaptive filtering algorithm based on the polarization covariance matrix is ​​selected to further suppress speckle noise while preserving edge structure and texture details.

[0082] (8) Terrain-Correction: Project the image from the slant distance / ground distance coordinate system to the geographic coordinate system, combine it with the DEM to realize geometric distortion correction, and output an orthophoto with real geographic coordinates, which is convenient for overlaying with other GIS data, spatial analysis and mapping output.

[0083] Within the forest area, the forest canopy is approximately a cloud layer composed of randomly oriented cylindrical scatterers. Ground scattering may originate from secondary scattering by a pair of orthogonal surfaces with different dielectric constants (e.g., the ground and the tree trunk), or from Bragg scattering by a slightly rough surface. It can be viewed as a scattering process of electromagnetic waves penetrating a layer of vertically oriented scatterers (FREEMAN A. Fitting a Two-Component Scattering Model to Polarimetric SAR DataFrom Forests [J]. IEEE Transactions on Geoscience and Remote Sensing, 2007,45(8): 2583–2592.).

[0084] A study proposed a polarimetric target decomposition method based on the Stokes polarimetric scattering eigenvector for dual-polarized Sentinel-1 SAR data (MASCOLO L, CLOUDE SR, LOPEZ-SANCHEZ J M. Model-Based Decomposition of Dual-Pol SAR Data: Application to Sentinel-1 [J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1–19.). The Stokes vector of the scattered wave is completely determined by the C2 matrix, specifically expressed as:

[0085] (1)

[0086] Stokes dual-polarization target decomposition utilizes the polarization Stokes matrix to decompose polarized scattered waves into α. 𝑣 and 𝑚 s Two polarization parameters represent the scattering components of the forest canopy and the dihedral angle scattering from the ground surface, respectively. The specific relationships are as follows:

[0087] (2)

[0088] in and The model parameters for surface scattering can be obtained using Stokes vectors:

[0089] .

[0090] Figure 3 This demonstrates the m-values ​​obtained after polarization decomposition from stitched Sentinel-1 data within the California region in December 2024. v and m s Parametric results. Both decomposition parameters present spatially continuous and clear results. For example... Figure 4 As shown, correlation analysis indicates that among the selected variables, m v It showed the strongest positive correlation with the water content of live fuel (LFMC) (r=0.57), while m s The correlation with dead fuel moisture content (DFMC) was the best (r=0.32). Compared with traditional meteorological factors such as precipitation, temperature, wind speed, and radar polarization backscattering coefficients (VV, VH), m v and m s They show a more significant correlation with the corresponding fuel moisture content. This indicates that they can more effectively characterize the moisture dynamics of live and dead fuels, are significantly superior to traditional environmental factors in terms of feature importance, and can serve as core input variables for subsequent FMC inversion.

[0091] 2.2 Preliminary estimation of forest fuel water content (FMC) using the Water Cloud Model (WCM): First, the optimal parameter set of the Water Cloud Model (WCM) is obtained through a two-stage optimization strategy, and then the forest fuel water content (FMC) is obtained by inversion using the Particle Swarm Optimization (PSO) algorithm.

[0092] To simulate the radar scattering characteristics of vegetated surfaces, in 1978, EPW Attema and Fawwaz T. Ulaby proposed a semi-empirical vegetation backscattering model (ATTEMA EPW, ULABY F T. Vegetation modeled as a water cloud [J]. Radio Science, 1978, 13(2): 357–364.). This model treats the vegetation canopy as a "water cloud," effectively explaining the backscattering phenomenon of microwave radar on vegetation, and is called the water cloud model. Some studies have used bare soil scattering models to represent the surface scattering component of the water cloud model, introducing surface parameters. This allows for a better understanding of how radar signals interact with vegetation and the land surface (WANG L, QUAN X, HE B, et al. Assessment of the Dual Polarimetric Sentinel-1A Data for Forest Fuel Moisture Content Estimation [J]. Remote Sensing, 2019, 11(13).), and the model is represented as:

[0093] (3)

[0094] For total radar scattering, A and B are empirical coefficients of the water cloud model, and P and Q are empirical coefficients of the bare soil scattering model. The radar incident angle is given. Different polarization modes are substituted into the model, and... The model is further expressed as follows, using the target parameters (i.e., the surface DFMC in this study) for parameterization:

[0095] (4)

[0096] (5)

[0097] Substituting the canopy scattering and surface scattering obtained through polarization decomposition into the model, the vegetation scattering and surface scattering of the water cloud model are decoupled. The model can be further expressed as:

[0098] (6)

[0099] (7)

[0100] in, and The tables separately represent the backscattering of canopy VV and canopy VH. and These represent the backscattering of VV and VH from the Earth's surface, respectively.

[0101] To establish the backscattering coefficient To determine the direct relationship with the fuel drought parameter DFMC, the intermediate variable LFMC needs to be eliminated. First, solve for LFMC from equation (9), take the natural logarithm of both sides of the equation and rearrange, then combine it with equation (10) to obtain equation (11):

[0102] (8)

[0103] 2.2.1 Parameter Optimization and FMC Solution

[0104] Combine equations (7) and (8) to establish the objective function for optimizing canopy parameters:

[0105] (9)

[0106] in, Let be the parameter vector to be optimized. Let N be the optimal parameter vector, and N be the number of observed samples. The VV polarization backscattering of the canopy calculated by the model, This refers to the actual observed VV polarization backscattering of the canopy. The canopy VH polarization backscattering calculated by the model, This is the actual observed VH polarization backscattering of the canopy.

[0107] Based on equation (11), establish the objective function for optimizing surface parameters:

[0108] (10)

[0109] in, Let be the parameter vector to be optimized. Let N be the optimal parameter vector, and N be the number of observed samples. The surface VH polarization backscattering is calculated by the model. This refers to the actual observed VH polarization backscattering at the Earth's surface.

[0110] To overcome the tendency of traditional optimization methods to get trapped in local optima, a two-stage hybrid optimization strategy is designed. The first stage is a Monte Carlo method pre-search stage, where 5000 sets of random number combinations are generated for both the canopy and the surface within the interval (0, 10], with 4 numbers per set for the canopy and 6 numbers per set for the surface. These random parameter combinations are used to solve the objective function, and the optimal parameter combination is recorded. The second stage is a Levenberg-Marquardt (LM) fine-tuning optimization. Since the LM algorithm is sensitive to initial values, this study selects the optimal parameter combination from the pre-search stage and uses a trust region adjustment strategy to solve the nonlinear least squares problem to obtain the optimal model parameters.

[0111] Based on equations (12) and (13), objective functions for univariate optimization to solve for the canopy LFMC and the surface DFMC are established respectively:

[0112] (11)

[0113] (12)

[0114] in, and Let LFMC and DFMC be the optimal canopy and surface combustible water content, respectively, and N be the number of observation samples. The efficient solution for both types of LFMC is achieved using the Particle Swarm Optimization (PSO) algorithm.

[0115] In this embodiment, the accuracy and time consumption of PSO, Genetic Algorithm (GA), and Simulated Annealing Algorithm (SA) in solving the FMC in univariate optimization were compared. Each algorithm was set with the same convergence criterion and optimized on the same 80 validation samples. The root mean square error (RMSE) of each optimization algorithm in solving the two types of FMC and the average time of each algorithm running the same 80 samples 10 times were compared to verify the adaptability of PSO in this method.

[0116] Figure 5This paper compares the performance of three optimization algorithms—PSO, GA, and SA—in solving the FMC problem. In LFMC estimation, PSO achieved the lowest RMSE (19.89); among the three algorithms, SA had the highest RMSE (20.10); GA achieved the same RMSE as PSO, but had the longest computation time (0.64 seconds). In DFMC estimation, PSO again performed best with the lowest RMSE (2.47) and the shortest computation time (0.21 seconds); GA achieved the same RMSE as PSO, but had the longest computation time (0.65 seconds); SA had an RMSE of 2.49 and a computation time of 0.22 seconds, inferior to PSO in both accuracy and efficiency. A comprehensive comparison of the three algorithms in both types of FMC estimation shows that PSO has a significant advantage in both accuracy and computational efficiency, highlighting its applicability in this study. Although GA can guarantee estimation accuracy to a certain extent, its long computation time makes it difficult to apply to large-scale FMC estimation; SA, while having computational efficiency close to PSO, has the lowest accuracy. Taking into account both estimation accuracy and efficiency, PSO is selected as the optimization algorithm for FMC inversion in the water cloud model (WCM) of this embodiment.

[0117] 2.2.2 Results of Preliminary Inversion of Canopy LFMC and Surface DFMC using the Water Cloud Model

[0118] Figure 6 The relationship between the estimated LFMC and DFMC based on the decomposed Sentinel-1 data and the observed values ​​is presented, where the optimal empirical coefficient is introduced into the WCM (see Table 1). The results show that the model has a moderate explanatory power for both types of FMC: LFMC... With a value of 0.62, it can explain approximately 62% of the data variance; DFMC's The variance was 0.52, explaining approximately 52% of the variance. The results also showed that when WCM was applied to the decomposed SAR data, it performed better than DFMC in estimating LFMC, indicating a more complex surface combustible condition, and that semi-empirical models like WCM have limitations in characterizing this complexity. Regarding error metrics, the RMSE for LFMC was 19.89 (unit: LFMC, %), and the RMSE for DFMC was 2.47 (unit: DFMC, %). Although the absolute RMSE of LFMC was significantly higher than that of DFMC, the NRMSE of LFMC (17.84%) was significantly lower than that of DFMC (31.17%) because the value range of DFMC (0–50%) was much narrower than that of LFMC (0–300%). Overall, WCM exhibited some systematic bias in estimating both types of FMC, indicating that the model still has room for further optimization.

[0119] Table 1. Optimal empirical coefficients for canopy LFMC and surface DFMC retrieved based on water cloud model

[0120] 2.2.3 Estimating LFMC and DFMC using the method of this embodiment

[0121] Figure 7 The FMC estimation results obtained using the method of this embodiment are shown; that is, the FMC estimate after machine learning error correction based on the initial WCM estimate. In the estimation of LFMC and DFMC, the method of this embodiment achieved R-values ​​of 0.75 and 0.72, respectively. 2 The value indicates that the model has a strong explanatory power for both types of FMC, and its explanatory power for DFMC is comparable to that for LFMC.

[0122] 2.3 Step 3) Error correction: Using SAR data, radar incident angle, slope and aspect as inputs, a machine learning model is constructed to perform nonlinear correction on the error in the initial forest fuel moisture content (FMC) estimation results obtained in step 2).

[0123] During Sentinel-1 data preprocessing, terrain smoothing is performed to mitigate the impact of terrain on synthetic aperture radar (SAR) signals (DOSTALOVA A, NAVACCHI C, GREIMEISTER-PFEIL I, et al. The effect of radiometric terrain flattening on SAR-based forest mapping and classification [J]. Remote Sensing Letters, 2022, 13(9): 855–864.). However, this processing cannot completely eliminate the influence of terrain. Furthermore, vegetation cover is another key factor affecting SAR signals. Due to the complexity of SAR scattering mechanisms, the aforementioned effects are difficult to describe effectively through physical modeling, leading to certain errors in FMC estimation. To further improve estimation accuracy, this embodiment introduces a machine learning model to correct the FMC estimation errors derived from the physical model, while ensuring adherence to physical constraints to a certain extent. To address the influence of terrain and vegetation, this study uses slope, aspect, and leaf area index (LAI) as additional input features of the model. The dataset is divided into training and testing sets. 60% of the samples are randomly selected for training, and the remaining 40% are reserved for testing. The model's input matrix is ​​as follows:

[0124] (13)

[0125] Where VV and VH represent backscattering with different polarization modes. The radar incident angle is denoted as y. The model output target y is the difference between the FMC derived from the semi-empirical model and the measured FMC.

[0126] (14)

[0127] To provide a clear comparison of the improvements made by machine learning, the dataset will not be re-split in this section. Furthermore, this study compares the error correction of two types of FMCs retrieved by the semi-empirical model using four machine learning models to verify the adaptability of RF in correcting errors in canopy LFMCs and MLP in correcting errors in surface DFMCs. These four models are RF, MLP, Gradient Boosting Decision Tree (GBDT), and Support Vector Regression (SVR).

[0128] 2.3.1 Comparison of the improvement in inversion accuracy by the four machine learning models

[0129] Tables 2 and 3 show the inversion accuracy of four machine learning models in LFMC and DFMC, as well as the 95% bootstrap confidence interval (CI). The results show that there are significant differences in the best models for these two fuel types: RF performs best in LFMC inversion, achieving the highest R² value (0.75, 95% CI: [0.67, 0.85]) and the lowest root mean square error (RMSE) (16.31%, 95% CI: [14.57, 17.94]%) and the lowest root mean square error (NRMSE) (14.62%, 95% CI: [8.80%, 14.91]%); while the multilayer perceptron (MLP) is more suitable for DFMC inversion, with an R² of 0.72 (95% CI: [0.63, 0.84]) and a reduced root mean square error (RMSE) of 1.89% (95% CI: [1.61, 2.15]%). This difference stems from the distinctly different hydrological response mechanisms of live fuels and dead fuels: LFMC, driven by meteorological factors, exhibits high variability and extreme fluctuations, while the ensemble averaging of RF effectively mitigates the impact of outliers; on the other hand, DFMC responds more slowly and has a lower baseline, making the nonlinear fitting capability of MLP more suitable for its gradual characteristics. Uncertainty analysis shows that the root mean square error (RMSE) confidence intervals for both fuel types remain narrow (width ≤ 1.0%), indicating high stability of the absolute prediction error; while the R² interval is relatively wide (range 0.20~0.35), mainly reflecting the spatial heterogeneity of fuel moisture content and sampling fluctuations due to the limited number of validation samples.

[0130] Table 2. Comparison of accuracy improvements of four machine learning models for canopy LFMC

[0131] Table 3. Comparison of the accuracy improvement of four machine learning methods for surface DFMC

[0132] 2.3.2, Inversion Mapping of LFMC and DFMC

[0133] Figure 8 and Figure 9The monthly average LFMC and monthly average DFMC of California in 2024, estimated using the method of this embodiment, are presented respectively. The western coastal region of California has a Mediterranean climate, with precipitation concentrated in winter and spring. The results show that from January to April, the LFMC of western California is in a relatively wet state. In May, the area with high LFMC values ​​in western California is significantly reduced compared to the previous four months. The eastern Sierra Nevada region of California has a highland climate. In winter and spring, it is controlled by the westerly winds, and precipitation is concentrated on the windward slopes on the west, forming a "rain shadow effect" (MULCH A, SARNA-WOJCICKI AM, PERKINS ME, etal. A Miocene to Pleistocene climate and elevation record of the Sierra Nevada (California) [J]. Proceedings of the National Academy of Sciences, 2008, 105(19): 6819–6824.). It can be seen that the LFMC during winter and spring forms a wet-dry boundary line along the direction of the Sierra Nevada. However, the moisture content of living fuels is not solely related to precipitation. Living trees acquire water from the environment in multiple ways (BERRY ZC, EMERY NC, GOTSCH SG, et al. Foliar water uptake: Processes, pathways, and integration into plant water budgets [J]. Plant, Cell & Environment, 2019, 42(2): 410–423.). Frequent dense fog in summer on the west coast of California provides plants with an important non-rainfall water source. Living trees in this region can directly absorb moisture from the fog through their canopy leaves (BAGUSKAS SA, OLIPHANT AJ, CLEMESHA RES, et al. Water and Light-Use Efficiency Are Enhanced Under Summer Coastal Fog in a California Agricultural System [J]. Journal of Geophysical Research: Biogeosciences, 2021, 126(5):e2020JG006193.). Therefore, the LFMC in western California did not show a significant decline from June to August due to reduced rainfall.The monthly average changes in DFMC and LFMC are not entirely consistent. Although both types of fuels are affected by environmental factors, dead surface fuels are inactive, and their composition, characteristics, and response mechanisms to environmental changes differ from those of live fuels. Therefore, their moisture content changes are not consistent (LEWIS CHM, LITTLE K, GRAHAM LJ, et al. Diurnal fuelmoisture content variations of live and dead Calluna vegetation in atemperate peatland [J]. Scientific Reports, 2024, 14(1): 4815.). The results show that DFMC did not remain relatively moist due to dense fog from June to August, which increased the risk of forest fires. This underscores the necessity of simultaneously retrieving these two types of FMC.

[0134] This embodiment also provides a SAR remote sensing-based forest canopy and surface combustible moisture content inversion device, including a processor. When the processor executes the program, it implements the above-mentioned SAR remote sensing-based forest canopy and surface combustible moisture content inversion method.

[0135] This embodiment also provides a computer-readable storage medium storing a computer program, which, when run by a processor, executes the steps of the above-described method for inverting the moisture content of forest canopy and surface combustibles based on SAR remote sensing.

[0136] This embodiment presents a SAR remote sensing-based method for inverting forest canopy and surface combustible water content, achieving simultaneous estimation of LFMC and DFMC, effectively addressing the long-standing challenge of vertical stratification estimation of FMC. Validation using a comprehensive ground-based dataset collected across California demonstrates the framework's robustness, with determination coefficients (R²) of 0.75 and 0.72 for LFMC and DFMC, respectively. The ensemble method proposed in this study achieves three key breakthroughs:

[0137] (1) Dual-parameter synchronous estimation capability: By effectively separating the canopy and surface scattering components through Stokes decomposition, the method in this embodiment only requires a single SAR image to simultaneously invert LFMC and DFMC. This breakthrough provides a reliable technical path for generating spatiotemporally consistent FMC distribution maps, significantly improving the accuracy and timeliness of dynamic wildfire risk assessment.

[0138] (2) Advantages of innovative hybrid modeling: By combining semi-empirical models with machine learning error correction, the framework retains the constraints of physical mechanisms while introducing nonlinear error compensation capabilities. This hybrid modeling strategy not only maintains the physical interpretability of the model but also significantly improves the accuracy and stability of parameter estimation.

[0139] (3) Good potential for operational application: The method is highly compatible with the existing Sentinel-1 satellite data system, and the framework adopts a modular architecture, which has strong adaptability and can be widely applied to various forest types and geographical environments.

Claims

1. A method for inverting the moisture content of forest canopy and surface combustibles based on SAR remote sensing, characterized in that, Includes the following steps: 1) Polarization decomposition: Using preprocessed Sentinel-1 IW SLC data, canopy feature parameters are extracted through Stokes decomposition. and surface characteristic parameters ; 2) Preliminary estimation of forest combustible moisture content using the water cloud model: First, the optimal parameter set of the water cloud model is obtained through a two-stage optimization strategy, and then the forest combustible moisture content is obtained by inversion using the particle swarm optimization algorithm. 3) Error correction: Using SAR data, radar incident angle, slope and aspect as inputs, a machine learning model is constructed to perform nonlinear correction on the error in the initial forest combustible moisture content estimation results obtained in step 2).

2. The method for inverting forest canopy and surface combustible moisture content based on SAR remote sensing according to claim 1, characterized in that: In step 1), canopy feature parameters are extracted using Stokes decomposition. and surface characteristic parameters It uses the polarization Stokes matrix to decompose the scattered wave into and Two polarization parameters represent the scattering components of the forest canopy and the dihedral scattering from the ground surface, respectively; their specific relationships are as follows: ； in and The model parameters for surface scattering can be obtained using Stokes vectors; 。 3. The method for inverting forest canopy and surface combustible moisture content based on SAR remote sensing according to claim 2, characterized in that: The Stokes vector of the scattered wave is entirely determined by the C2 matrix, specifically expressed as: 。 4. The method for inverting forest canopy and surface combustible moisture content based on SAR remote sensing according to claim 1, characterized in that: The preprocessing in step 1) includes one or more of the following: orbit correction, radiometric calibration, pulse band splicing, polarization matrix generation, multi-view, terrain smoothing, polarization spot filtering, and terrain correction.

5. The method for inverting forest canopy and surface combustible moisture content based on SAR remote sensing according to claim 1, characterized in that: The optimal parameter set of the water cloud model in step 2) is obtained through a two-stage optimization strategy, which includes a Monte Carlo pre-search stage and a least squares fine optimization stage. The Monte Carlo method pre-search stage involves generating 5000 sets of random number combinations in the interval (0, 10) for both the canopy and the ground surface. Each set of the canopy contains 4 numbers, and each set of the ground surface contains 6 numbers. These random parameter combinations are used as random parameter combinations to optimize the parameters. The objective function is solved using these 5000 sets of random parameter combinations, and the optimal parameter combination is recorded. The least squares fine optimization stage involves selecting the optimal parameter combination from the pre-search stage, solving the nonlinear least squares problem through a trust region adjustment strategy, and obtaining the optimal model parameters.

6. The method for inverting forest canopy and surface combustible moisture content based on SAR remote sensing according to claim 1, characterized in that: The forest combustible moisture content in step 2) includes the forest canopy combustible moisture content and the forest surface combustible moisture content; ； ； in, and LFMC and DFMC represent the optimal canopy and surface combustible water content, respectively, and N represents the number of observation samples.

7. The method for inverting forest canopy and surface combustible moisture content based on SAR remote sensing according to claim 1, characterized in that: The input matrix of the machine learning model in step 3) is: ； Where VV and VH represent backscattering with different polarization modes. The radar incident angle; The model output target y is the difference between the FMC derived from the semi-empirical model inversion and the measured FMC. 。 8. A device for retrieving forest canopy and surface combustible moisture content based on SAR remote sensing, characterized in that: Includes a processor, which, when executing a program, implements the method described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the method for retrieving forest canopy and surface combustible moisture content based on SAR remote sensing as described in any one of claims 1-6.

Images