Combustible moisture content inversion method based on mechanism model and artificial neural network and application

Abstract

The invention relates to a combustible moisture content inversion method based on a mechanism model and an artificial neural network and application. The combustible moisture content inversion method comprises the following steps: S1, constructing a canopy reflectivity simulation data set; s11, simulating the canopy reflectivity of the grassland vegetation by using a PROSAIL model; s12, simulating a non-uniform vegetation canopy reflectivity by using a 4-Scale geometrical optical model; s13, excluding unreasonable data by using the function relationship among the EWT, the DMC, the Cab and the LAI value; s14, eliminating unreasonable data based on Gaussian white noise; s15, acquiring simulated canopy reflectivity data corresponding to the MODIS wavelength range; s2, constructing an FMC inversion model; s21, training a mechanism model; S22, optimizing the mechanism model to obtain an FMC inversion model; and S3, realizing FMC inversion. The method has the characteristics of high precision, self-learning, rapidness, high efficiency and the like.

Description

Technical Field

[0001] This application relates to a method and application for vegetation remote sensing inversion parameters, particularly a method and application for inverting the moisture content of combustibles based on a mechanism model and artificial neural network, which is mainly applicable to the estimation of the moisture content of vegetation combustibles. Background Technology

[0002] Forest fires are random in nature, resulting from the interplay of multiple factors, but they also exhibit inherent patterns. The well-known fire triangle model identifies climate, topography, and fuel composition as three crucial indicators for assessing fire risk. Fuel composition information includes fuel moisture content (FMC), fuel load, and fuel type. Among these, FMC is closely related to the occurrence of forest fires, determining the ease with which a forest can be ignited, thus influencing fire behavior and serving as a key indicator in forest fire risk assessment and fire spread rate calculation.

[0003] Traditional forest fuel load (FMC) measurement involves field sampling in the study area, weighing the samples (wet weight), drying them, and weighing them again (dry weight). The ratio of moisture content (wet weight - dry weight) to dry weight is then calculated to obtain the FMC of the sample plot. Field sampling is time-consuming and labor-intensive, yielding data from only a limited number of points. Its advantage is high accuracy, and the observational data can be used to validate data products generated by other research methods. Another widely used method is the meteorological factor regression method, which performs multivariate statistical regression on the collected sample plot fuel load and related meteorological factor data to establish a regression estimation model. This method is simple to model; however, vegetation fuel load is affected by both meteorological conditions and the vegetation's own life activities, so models built solely using meteorological factors have low accuracy and lack universality. Compared to traditional field FMC measurements, remote sensing technology can provide large-scale, spatiotemporally continuous image data, and the near-infrared and shortwave infrared bands are sensitive to vegetation canopy moisture content, making it the primary technology for near-real-time, large-area, and spatiotemporally continuous monitoring of fuel load. The FMC calculation formula is as follows:

[0004] In formula (1), EWT represents Equivalent Water Thicknes, and DMC represents Dry Matter Content, both in g / cm³. 2 .

[0005] Currently, methods for estimating vegetation canopy MC using remote sensing technology can be divided into three categories: (i) estimation methods based on empirical statistical models; these methods directly utilize empirical statistical models such as multiple linear regression to remotely estimate the vegetation canopy MC, aiming to find the linear (or nonlinear) relationship between the vegetation canopy MC and multispectral reflectance data or relevant typical vegetation indices (such as Leaf Area Index (LAI), Normalized Difference Vegetation Index (NDVI), Normalized Difference Infrared Index (NDII), and Enhanced Vegetation Index (EVI)). (ii) inversion methods based on physical models (canopy reflectance models); these methods simulate vegetation canopy reflectance using physical models while inverting key biochemical parameters, including Equivalent Water Thicknes (EWT) and Dry Matter Content (DMC), and then obtain the MC based on their ratio. (iii) Inversion method based on neural networks; deep learning is a branch of machine learning and is included within machine learning. Deep learning models can approximate the complex nonlinear relationship between various biological and geophysical parameters and remote sensing data through multi-layer learning. Based on the construction and training of artificial neural networks, a model is established by using artificial neural network (ANN) algorithms to estimate vegetation biochemical parameters EWT and DMC related to FMC, thereby obtaining a model for inverting FMC. Summary of the Invention

[0006] The technical problem solved by this application is to provide a high-precision and easy-to-use mechanistic model and ANN method, which uses a radiative transfer physics model to generate high-quality training sample data and couples it with a machine learning-style mechanistic model, providing important support for forest fire risk assessment and ecological monitoring.

[0007] The main principle of this application is as follows: First, a massive amount of representative MODIS canopy reflectance data for different vegetation types is generated by simulating the PROSAIL model based on radiative transfer theory and the 4-Scale model based on geometrical optics theory. Second, a mechanistic model considering vegetation type is constructed based on the simulated canopy reflectance data. Third, the trained mechanistic model (inversion model) is deployed on a global vegetation status monitoring system (geospatial data monitoring system, such as a GEE platform) to perform FMC inversion, obtaining spatiotemporally continuously changing FMC inversion data on a global scale. Finally, publicly available ground-based measured data are used to verify the simulation results and evaluate the applicability and sensitivity of the inversion model.

[0008] The technical solution adopted by this application to solve the above-mentioned technical problems includes: a method for inverting the moisture content of combustibles based on a mechanism model and an artificial neural network, which mainly includes the following steps: S1: Construct a simulated dataset of canopy reflectance based on the IGBP classification system; S11: Simulate the reflectance of grassland vegetation canopy using the PROSAIL model (which includes the PROSPECT model and the SAIL model), with a simulation band of 400-2500 nm and a spectral resolution of 1 nm. S12: The 4-Scale geometric optics model was used to simulate the canopy reflectance of non-uniform vegetation (composed of shrubs, broad-leaved forests and coniferous forests). The input parameters of the 4-Scale geometric optics model included the leaf reflectance and transmittance obtained from the PROSPECT model. The simulation band was 400-2500nm and the spectral resolution was 1nm. S13: Using the functional relationship between the equivalent water thickness, dry matter content, chlorophyll content, and leaf area index (LAI) value of leaves as constraints to eliminate unreasonable data, the range of variation of equivalent water thickness, dry matter content, and chlorophyll content is limited as follows:

[0009]

[0010] In formulas (2) and (3), V 𝑚𝑖𝑛 (LAI), V max (LAI) represents one of the parameters—equivalent water thickness, dry matter content, and chlorophyll content—as LAI varies, V 𝑚𝑖𝑛 V 𝑚𝑎𝑥 These represent the minimum and maximum values ​​of the parameter when LAI reaches its minimum value, respectively. min (LAI 𝑚𝑎𝑥 V 𝑚𝑎𝑥 (LAI 𝑚𝑎𝑥) represent the maximum and minimum values ​​of this parameter when LAI is at its maximum value. The parameter values ​​corresponding to different vegetation types are based on the data in Table 5. S14: Based on Gaussian white noise, additive and multiplicative uncertainties, including band-correlated and band-independent uncertainties, were added to the simulated reflectivity data.

[0011] In formula (4), where and These represent the reflectivity values ​​of the λ-band simulated by the radiative transfer model and the reflectivity values ​​after adding noise, respectively. Represents a normal distribution. It is a relative uncertainty that applies to all wavebands. This refers to the relative uncertainty applied to the λ-band. It is an absolute uncertainty that applies to all wavebands. This refers to the absolute uncertainty applied to the λ-band. = = 0.01, = = 4%; S15: The processed simulated reflectance data is transformed according to the MODIS spectral response function to obtain simulated reflectance data for 7 bands corresponding to the MODIS wavelength range. The MODIS bands are described below:

[0012] S16: The vegetation indices related to EWT and leaf DMC are calculated based on the canopy reflectance of the seven bands of MODIS (expressed as Band1, Band2, Band3, Band4, Band5, Band6, and Band7, respectively). S2: Construct an FMC inversion model based on the IGBP classification system; S21: The simulated datasets for the four vegetation types are divided into two parts. One part is used for neural network training, and the other part is used to test the performance of the mechanistic model. The simulated datasets contain parameters from band 1 to band 7, vegetation indices, and leaf area indices. The mechanistic model is expressed as follows: In formula (1), FMC is the combustible water content of the vegetation canopy, EWT is the equivalent water thickness of the vegetation canopy, and DMC is the dry matter content of the vegetation canopy. S22: Train the neural network until the output of the mechanism model meets the requirements. At this point, the mechanism model is the FMC inversion model. S3: Achieve FMC inversion through a global vegetation status monitoring system based on the IGBP classification system.

[0013] Furthermore, this application may also include step S4. S4: Verify the inversion results and evaluate the performance of the FMC inversion model.

[0014] Step S22 of this application includes the following steps: S221 Input Layer Optimization: To determine the optimal input parameters, with the output layer, hidden layer, and other parameters of the mechanistic model fixed, the optimal band combination is determined by comparing the importance of each parameter. Parameter selection utilizes a sequential forward search algorithm to select the subset that contributes most to the performance of the mechanistic model from a given feature set. During parameter selection, the maximum coefficient of determination R0 for the EWT and DMC parameters in the mechanistic model results is chosen. 2 The optimal band combination with the minimum normalized root mean square error (nRMSE) serves as the basis for the next parameter increase. The optimal combination is the input parameter combination when the parameter optimization accuracy meets the requirements or is difficult to improve. Through parameter optimization, the mechanism model with the best EWT and DMC inversion results for the four vegetation types can be obtained.

[0015] S222 Output Layer Optimization: With the input layer, hidden layer, and other model parameters fixed, the performance of the mechanistic model was compared when EWT, DMC, Car, Cab, and Ant were used as output variables, when EWT and DMC were used as co-output variables, and when EWT and DMC were used as single output variables. The comparison results show that the mechanistic model performs better overall (referring to the combination of accuracy and computational efficiency) when EWT and DMC are used as co-output variables.

[0016] S223 Hidden Layer Optimization: Hidden layer parameters were tuned by continuously adjusting the activation function, learning rate, and optimizer parameters of the mechanistic model to obtain the optimal mechanistic model parameters. The number of neurons in the three hidden layers were 64, 128, and 64, respectively. The loss function was mean squared error. The activation function for each hidden layer was the Tanh function, the activation function for the output layer was the Linear function, the optimization function for the output layer was Adam, and the learning rate for the output layer was 0.005.

[0017] In step S22 of this application, vegetation index parameters can be added during input layer optimization to improve accuracy. As a special case, the vegetation index parameters applied in this application include normalized differential infrared index, normalized vegetation index, enhanced vegetation index, water stress index, normalized differential water index, simple ratio water index, shortwave infrared water stress index, normalized spectral angle index, and similarity water index.

[0018] S3: FMC inversion is achieved through the GEE platform; S31: Preprocessing of remote sensing data from MCD43A4, MCD12Q1 and MCD15A3H; S32: Deploy the FMC inversion model to the GEE platform, obtain the inversion model parameters to perform EWT and DMC inversion, and limit their numerical range to the maximum and minimum range of the simulated reflectance data. Then, obtain the FMC result based on EWT / DMC and control the FMC value within the normal range of 0%-250%.

[0019] The global vegetation status monitoring system based on the IGBP classification system uses Google Earth Engine.

[0020] The technical solution adopted in this application to solve the above-mentioned technical problems also includes: further application of the above-mentioned combustible material moisture content inversion method: S5: Acquisition and representation of the spatial distribution characteristics of global FMC.

[0021] S6: Acquisition and representation of global FMC temporal distribution characteristics.

[0022] Compared with existing products, this application has the following advantages and effects: 1. The geometrical optical model considers factors such as vegetation structure and spectral reflectance characteristics, and models based on physical principles, which improves the interpretability and accuracy of the mechanistic model; 2. The mechanistic model can automatically learn features from complex remote sensing data, capture nonlinear relationships, and provide high-precision predictions; 3. By combining global remote sensing data with neural networks and geometrical optical models, combustible material moisture content can be efficiently retrieved over a large area, without regional limitations; 4. The neural network can be trained on large datasets, quickly process high-dimensional, multi-source remote sensing data, and can be updated in real time to track changes in combustible material moisture content globally; 5. Combustible material moisture content is an important parameter for assessing fire risk, climate change, and ecosystem changes. This model provides a new and efficient method for global fire risk assessment and ecological disaster prediction. Attached Figure Description

[0023] Figure 1 The distribution patterns of EWT, DMC, and Cab are shown in relation to LAI. Formulas (2) and (3) demonstrate the changing characteristics and trends of EWT, DMC, and Cab when considering LAI variations. For example, in natural environments, when the leaf area index is high, other ecological parameters rarely have low levels.

[0024] Figure 2This document presents a schematic diagram of the 2021 reclassification (grassland, broadleaf forest, coniferous forest, and shrub) of land cover types in MODIS data using the MCD12Q1 classification system based on the IGBP classification scheme, and displays FMC validation site information. The MCD12Q1 product provides global land cover maps with an annual time step and 500-meter spatial resolution from 2001 to the present. This application uses the IGBP product classification system to divide global vegetation into four categories (grassland, shrub, coniferous forest, and broadleaf forest) and displays them in different colors on the map. This application statistically analyzes Globe-LFMC data, summarizing site data from multiple countries including Australia and Italy from 2001 to 2018, covering thousands of FMC field measurements of various vegetation types such as grassland, shrub, and forest. Through temporal consistency and spatial homogeneity tests of the site data, outlier data points in consecutive observations were removed using temporal rules, and spatial heterogeneity within pixels was determined based on the coefficient of variation. Finally, validation points were selected as shown by the red dots in the figure.

[0025] Figure 3 The technical flowchart for retrieving global canopy reflectance (FMC) from MODIS data using mechanistic models and artificial neural network methods consists of three steps. First, a massive amount of representative MODIS canopy reflectance data for different vegetation types is generated using the PROSAIL model based on radiative transfer theory and the 4-Scale model based on geometrical optics theory. Second, a mechanistic model considering vegetation type is constructed and trained based on the simulated canopy reflectance data to obtain the trained inversion model. Finally, the trained inversion model is deployed on the GEE platform to perform FMC inversion, obtaining spatiotemporally continuously varying FMC simulation (results) data globally.

[0026] Figure 4 The diagram shows the changes in R² and nRMSE accuracy of the mechanistic models for different vegetation types on the test set as the input parameters increase (where Blue represents the blue band, Red represents the red band, Green represents the green band, and SWIR1, SWIR2, and SWIR3 represent the MODIS bands with wavelength ranges of 1230-1250 nm, 1628-1652 nm, and 2105-2155 nm, respectively). For all vegetation types, EWT's inversion accuracy is higher than DMC's. Grassland ( Figure 4 The EWT and DMC inversion accuracy of broadleaf forests (A) is higher than that of other vegetation types, and the accuracy reaches saturation after gradually increasing the last few bands; the EWT inversion accuracy of broadleaf forests, coniferous forests and shrubs is significantly higher than that of DMC, especially in broadleaf forests (A). Figure 4 B) is more obvious.

[0027] Figure 5This paper illustrates the accuracy variations of mechanistic models trained on the test set for broadleaf forests under different output parameters. The main comparisons focus on multiple output variables (EWT, DMC, carotenoid content (Car), chlorophyll content (Cab), and anthocyanin content (Ant)), co-output variables (EWT and DMC), and individual outputs (EWT and DMC). The accuracy is lower when considering multiple outputs (EWT, DMC, Car, Cab, and Ant), while better when considering two co-outputs. Although the accuracy is slightly higher when EWT and DMC are used as outputs individually, this requires two separate mechanistic models corresponding to EWT and DMC, increasing the number of models and reducing computational efficiency.

[0028] Figure 6 The comparison and verification of the predicted FMC and the measured FMC in this application are shown. A total of 1619 sets of data were used for verification. Figure 6 Based on the distribution of validation sample points, the FMC value range was divided into 8 intervals. All samples were assigned to an 8×8 grid, with the mean of the measured and predicted values ​​of each sample in the grid used as the center coordinates, and the normalized standard deviation of the predicted values ​​(SDn=std / mean) used as the radius. Colors were assigned proportionally to the number of sample points, where 'n' represents the sample size. Validation results showed a correction coefficient of 0.70 and a correlation coefficient R0. 2 The significance level was 0.49, P < 0.0001, Bias = 8.16, MAE = 24.38, RMSE = 29.02, indicating significant validation. As shown in the figure, the sample points were mainly concentrated around 50%-100% of the FMC. The lower the measured and predicted FMC values, the larger the SDN (Survey Difference), and vice versa. The larger SDN values ​​were all distributed below the 1:1 line, indicating that the difference in prediction results was greater at low FMC values ​​than at high FMC values.

[0029] Figure 7 Histograms of the mean and frequency distributions of the global 500m resolution seasonal FMC spatial distribution for 2021-2022, obtained using the method of this application, are shown. (Spring: March to May; Summer: June to August; Autumn: September to November; Winter: December to February of the following year).

[0030] Figure 8 The global average FMC for four periods and four vegetation types from 2021 to 2022, calculated according to latitude, is shown using the method of this application.

[0031] Figure 9 The temporal evolution of monthly average FMC for different vegetation types in different temperature zones worldwide from 2010 to 2021 is shown using the method of this application. Detailed Implementation

[0032] Explanation of the main names in this application: Combustible material moisture content: Abbreviated as FMC. Since this application is based on a mechanistic model, the canopy FMC = canopy EWT / canopy DMC = (leaf EWT * LAI) / (leaf DMC * LAI) = leaf EWT / leaf DMC. Therefore, this application can use leaf FMC to simulate canopy FMC. Equivalent water thickness: Abbreviated as EWT, actually refers to the blade equivalent water thickness; Dry matter content: Abbreviated as DMC, but actually refers to the dry matter content of leaves.

[0033] A method for retrieving the moisture content of combustibles based on a mechanistic model and artificial neural networks mainly includes the following steps: S1: Construct a simulated canopy reflectance dataset for four vegetation types (grassland, shrubland, broadleaf forest, and coniferous forest reclassified based on the IGBP classification system); S11: The PROSAIL model (including the PROSPECT model and the SAIL model) is used to simulate the canopy reflectance of grassland vegetation, such as crops and grassland, with a one-dimensional horizontal continuous uniform distribution. The simulation band is 400-2500nm and the spectral resolution is 1nm. The specific parameter settings are shown in Tables 2 and 3. S12: Simulate the canopy reflectance of non-uniform vegetation (composed of shrubs, broad-leaved forests and coniferous forests) using a 4-Scale geometric optics model. The input parameters of the 4-Scale geometric optics model include the leaf reflectance and transmittance obtained from the PROSPECT model simulation in step S11. The simulation band is 400-2500nm, and the spectral resolution is 1nm. See Table 2 and Table 4 for the specific parameter settings. S13: The relationship between the ranges of EWT, DMC, and Cab and the LAI value is used as a constraint to avoid generating unlikely simulated spectra (eliminating meaningless parameter combinations and reducing computational complexity). It is only necessary to assume that the ranges of variation change linearly with LAI between Vmin(Vmax) and Vmin(LAImax)(Vmax(LAImax)). The relationship between the ranges of EWT, DMC, and Cab as a function of the LAI value is shown in the following formula:

[0034]

[0035] In formulas (2) and (3), V 𝑚𝑖𝑛 (LAI), V max (LAI) represents the minimum and maximum values ​​of one of the parameters EWT, DMC, and Cab as LAI changes, respectively, V 𝑚𝑖𝑛V 𝑚𝑎𝑥 These represent the minimum and maximum values ​​of the parameter when LAI reaches its minimum value, respectively, and V. min (LAI 𝑚𝑎𝑥 V 𝑚𝑎𝑥 (LAI 𝑚𝑎𝑥 ) represent the maximum and minimum values ​​of this parameter when LAI is at its maximum value. The parameter values ​​corresponding to different vegetation types are shown in Table 5. S14: Based on Gaussian white noise, additive and multiplicative uncertainties, including band-correlated and band-independent ones, were added to the simulated canopy reflectivity data.

[0036] In formula (4), where and These represent the reflectivity values ​​of the λ-band simulated by the radiative transfer model and the reflectivity values ​​after adding noise, respectively. This represents a normal distribution (mean equal to 0 and variance σ²). It is a relative uncertainty that applies to all wavebands. This refers to the relative uncertainty applied to the λ-band. It is an absolute uncertainty that applies to all wavebands. This refers to the absolute uncertainty applied to the λ-band. The absolute uncertainty and relative uncertainty values ​​used in this application are respectively... = = 0.01 and = = 4%; S15: The processed simulated reflectance data is resampled (data conversion) according to the MODIS spectral response function to obtain 7 bands corresponding to the MODIS wavelength range (the original simulated reflectance data is the reflectance at 400-2500 nm, the MODIS sensor divides the 400-2500 nm wavelength range into 7 bands: Band 1 (Red, 620–670 nm): Band 2 (NIR, 841–876 nm): Band 3 (Blue, 459–479 nm): Band 4 (Green, 545–565 nm): Band 5 (Red Edge, 1230–1250 nm): Band 6 (SWIR, 1628–1652 nm): Band 7 (SWIR, 1628–1652 nm): Band 8 (NIR, 1628–1652 nm): Band 9 (NIR, 1628–1652 nm): Band 10 (NIR, 1628–1652 nm): Band 11 (NIR, 1628–1652 nm): Band 12 ... (2105–2155nm)}, the average reflectance within this range is collected for each band as the reflectance of that band, so it is necessary to convert the simulated reflectance data within this range into simulated canopy reflectance data of a single value; S16: Based on the canopy reflectance of the seven bands of MODIS, vegetation indices related to EWT and DMC were calculated using the formula in Table 6. S2: Construct an FMC inversion model that takes into account vegetation type; S21: The simulated datasets for four vegetation types (seven reflectance data points corresponding to bands 1-7, vegetation indices, and leaf area indices) were randomly divided into two parts. 90% of the data was used for neural network training, and 10% of the data was used to test the performance of the mechanistic model after training. The mechanistic model is expressed as follows:

[0037] S22: Train and optimize the input, output, and hidden layers of the neural network until the output of the mechanistic model meets the requirements. S221 Input Layer Optimization: This mainly involves determining the optimal input parameters, considering bands and the vegetation indices shown in Table 6. With the output layer, hidden layers, and other model parameters fixed, the optimal band combination is determined by comparing the importance of each parameter. Parameter selection primarily utilizes the Sequential Forward Selection (SFS) algorithm, a feature selection algorithm, a type of greedy algorithm, used to select the subset that contributes most to model performance from a given feature set. During the process, the maximum coefficient of determination R0 for the EWT and DMC parameters in the model results is selected. 2 The band combination with the minimum normalized root mean square error (nRMSE) serves as the basis for the next parameter increase. The optimal combination is considered to achieve the required accuracy or when parameter increases result in minimal performance improvement but a significant increase in computational cost. The input parameter optimization process is detailed below. Figure 4 By optimizing the parameters, we can obtain the model with the best EWT and DMC inversion results for the four vegetation types.

[0038] S222 Output Layer Optimization: With the input layer, hidden layer, and other model parameters fixed, the performance is mainly compared when multiple output variables (EWT, DMC, Car, Cab, and Ant), co-output variables (EWT and DMC), and when EWT and DMC are used as single output variables respectively. Models for the four vegetation types were trained under different combinations of output parameters, and their inversion results are shown in [link to documentation]. Figure 5 . Figure 5This describes the accuracy variation of the broadleaf forest vegetation model under different output parameters. Accuracy is lower when considering multiple outputs (EWT, DMC, Car, Cab, and Ant), while accuracy is better when considering two co-outputs. Although the model accuracy is slightly higher when EWT and DMC are used as outputs individually than with co-outputs, this requires two separate mechanistic models corresponding to EWT and DMC respectively, increasing the number of models and reducing computational efficiency. Therefore, the co-output approach is used. The difference in accuracy here mainly stems from the number of output parameters. This principle also applies to other vegetation types; therefore, co-output parameter combinations are also used for coniferous forests, shrubs, and grasslands.

[0039] S223 Hidden Layer Optimization: The main focus in the hidden layer is parameter tuning. By continuously adjusting parameters such as the activation function, learning rate, and optimizer, the optimal model parameters are finally obtained: 3 hidden layers (with 64, 128, and 64 neurons respectively), the loss function is mean squared error (MSE), the activation function for each hidden layer is the Tanh function, the activation function for the output layer is the Linear function, the optimization function is Adam, and the learning rate is 0.005.

[0040] In step S22 of this application, during input layer optimization, indices can be added to improve accuracy, such as the normalized spectral angle index (NSAI), similarity water indices (SWI), and normalized difference infrared index (NDVI). The specific vegetation indices used in this application are shown in Table 6, and the impact of adding indices on the model is discussed in [Table 6]. Figure 4 .

[0041] S3: Achieve FMC inversion through the GEE (Google Earth Engine) platform; S31: Preprocessing of remote sensing data from MCD43A4, MCD12Q1 and MCD15A3H; S32: Deploy the FMC inversion model to the GEE platform, obtain the model parameters to perform EWT and DMC inversion, and limit their numerical range to the maximum and minimum range of the simulated reflectance data. Then, obtain the FMC result based on EWT / DMC and control the FMC value within the normal range of 0%-250%.

[0042] S4: Verify the inversion results and evaluate the performance of the FMC inversion model.

[0043] The main applications of this application include the following aspects: S5: Acquisition and representation of the spatial distribution characteristics of global FMC.

[0044] S6: Acquisition and representation of global FMC temporal distribution characteristics.

[0045] The FMC inversion model described in this application is highly adaptable and can be easily updated because it covers real-world vegetation types and simulates a large amount of canopy reflectance data. Generally, obtaining global FMC data is sufficient to obtain FMC information for the desired region.

[0046] This application uses the coefficient of determination (R²) 2 The performance of the model on the test set is evaluated using the normalized root mean square error (nRMSE%) and the correction coefficient (r), and the coefficient of determination (R²) is used. 2 The direct validation performance of the model inversion results is evaluated using the significance level (P), bias, mean absolute error (MAE), and root mean square error (RMSE). The final evaluation results of the inversion model are reflected in... Figure 4 The rightmost of the four smaller graphs contains the specific data.

[0047] The PROSAIL model described in this application is a combination of the SAIL canopy radiative transfer model and the PROSPECT leaf optical model. The PROSPECT leaf optical model simulates the reflectance spectra of leaves from different vegetation types by inputting leaf trait data, including leaf chlorophyll content, leaf dry matter content, leaf equivalent water thickness, leaf carotenoid content, leaf brown pigment content, leaf structural parameters, and anthocyanin content. The PROSAIL model describes the multiple scattering and directional reflection characteristics of light in the vegetation canopy using a physical parameterization method. Input parameters include solar zenith angle, observed zenith angle, relative azimuth angle, leaf area index, hotspot effect factor, soil factors, and leaf tilt angle distribution factor, as well as the leaf reflectance and transmittance simulated by the PROSPECT model. This application further corrects the leaf reflectance and transmittance of grassland vegetation simulated by the PROSPECT model from the perspective of factors affecting optical effects by adding solar zenith angle, observed zenith angle, relative azimuth angle, leaf area index, hotspot effect factor, soil factors, and leaf tilt angle distribution factor.

[0048] The 4-Sacle model described in this application is an advanced geometric optics model developed by Fan et al. based on the geometric optics model and combined with ray tracing technology. It has rich input parameters, covering plot geometric parameters (such as plot size, LAI, tree density, aspect, slope, solar and observational geometric information), tree structural parameters (such as tree diameter at breast height, crown size, height below the first branch, clump index, etc.), and the spectral reflectance characteristics of leaves and background. This application further corrects the leaf reflectance and transmittance of shrublands, broad-leaved forests, and coniferous forests simulated by the PROSPECT model from the perspective of factors affecting optical effects by adding plot geometric parameters, tree structural parameters, and the spectral reflectance characteristics of leaves and background, demonstrating strong specificity.

[0049] To ensure the representativeness of the simulated dataset, this application obtained the size range of normal vegetation traits in nature by consulting existing public literature, traversed these size ranges, and used the aforementioned PROSAIL model and 4-Scale model to generate a representative vegetation simulation dataset. The generated simulation dataset sample includes: canopy reflectance spectrum and corresponding leaf moisture content and dry matter content.

[0050] This application generated global FMC data products from 2010 to 2023. Due to the rigorous quality screening and control of the remote sensing data in the early stages, the effective data area of ​​the inversion results is relatively limited. Furthermore, the LAI data from MCD15A3H contains quality information, and after combining it with the quality data from MCD43A4, the effective pixels of the LAI data are further reduced. Therefore, this study designed two different models for the inclusion of LAI data in the model input variables. The remote sensing data used in this study include MCD43A4 reflectance data and its derived vegetation indices and LAI data, and the impact of these data on the models under quality screening and without quality screening was explored. Through comprehensive consideration of different data combinations and their quality processing, the final global FMC inversion results were obtained.

[0051] The following analysis of the inversion results further describes this application: Utilizing GEE's high-performance parallel computing platform, this application generated a dynamic map of seasonal global FMC changes from 2021 to 2022 using MODIS data. Figure 7The frequency distribution histogram corresponds to this. In spring, mid-to-high latitude regions of the Northern Hemisphere (such as Europe, North America, and East Asia) show higher FMC (Fresh Moisture Content), which is related to the start of vegetation growth and increased water content. The Southern Hemisphere, such as Australia and South Africa, shows lower FMC, indicating that the dry climate in these regions affects plant water content. In summer, FMC in the Northern Hemisphere peaks, especially in forested areas such as North America, Europe, and western Russia, with water content exceeding 120%, and green to light yellow indicating vigorous vegetation growth. FMC in the Southern Hemisphere shows a downward trend, especially in southern South America and parts of Australia, possibly because this season is a dry period with lower water content. In autumn, as the Northern Hemisphere enters autumn, FMC decreases significantly, particularly in mid-to-high latitude regions of North America, Europe, and Asia (such as northern China and Russia), as vegetation gradually enters dormancy and water content decreases. The Southern Hemisphere begins to recover FMC, especially in central South America and South Africa, where water content increases. During winter, the fuel moisture content (FMC) in the Northern Hemisphere drops to its lowest point, particularly in North America, Europe, and Russia, showing large areas of dark blue that represent dry or low-moisture winter vegetation. Parts of the Southern Hemisphere, such as South America, South Africa, and parts of Australia, show slight recovery, but the overall FMC remains low, indicating that the fuel moisture content of vegetation in these regions remains low.

[0052] In high-latitude regions, such as Canada, Russia, and Northern Europe, the highest FMC values ​​(approaching 120% or higher) are observed in summer, but drop sharply to lower levels in winter. This is typical of the vegetation cycle in high-latitude regions, significantly influenced by seasonal climate change. In mid-latitude regions of North America, Europe, Central Asia, and East Asia, seasonal variations in FMC are very pronounced. In spring and summer, FMC gradually increases in these regions, but declines rapidly in autumn and winter. Particularly in northern China and the western United States, FMC decreases significantly in autumn and winter, possibly related to increased forest fire risk.

[0053] In summary, seasonal variation is significant in the Northern Hemisphere, particularly in mid- to high-latitude regions where the free-floating mass (FMC) peaks in summer and declines rapidly in autumn and winter. Variation in the Southern Hemisphere is relatively mild, but parts of Australia, South America, and South Africa show significant FMC declines in summer, reflecting the effects of seasonal drought. The FMC in the tropics is relatively stable throughout the year, but noticeable seasonal variations are observed in parts of East Africa and Australia, reflecting localized climatic conditions.

[0054] Figure 8 for Figure 7Average FMC curves for different latitudes and vegetation types. Overall, the FMC curve for broadleaf forests exhibits the highest dynamics, followed by shrubland, coniferous forest, and forest curves. For all vegetation types, FMC values ​​are higher in summer than in other seasons and fluctuate more significantly, reflecting the high temperatures and rainfall of summer. FMC values ​​are more similar and change less in the other three seasons. Grassland ( Figure 8 A) In the Southern Hemisphere below the equator, the FMC (free flow rate) varies less, while in the Northern Hemisphere above 0°, the FMC is more distinct. Except for spring, the FMC does not fluctuate much in other seasons, but shows a significant increase in spring, and then begins to decline after 30°, stabilizing around 60°. (Broadleaf forests) Figure 8 B) The FMC value varies significantly across all latitudes, with larger fluctuations in summer and slight fluctuations in other seasons, but the FMC value remains relatively stable at around 100%; coniferous forests ( Figure 8 C) The FMC value shows a consistent trend throughout the four seasons. At -40°, the FMC value begins to decrease from 150% to 60%, with no data available near the equator, indicating that the equatorial region is unsuitable for coniferous vegetation growth. Subsequently, the FMC value begins to increase again after 40°; shrubland ( Figure 8 D) Its dimensional variation trend is not much different from that of coniferous forests, but its FMC value starts to decrease at -20° and starts to increase at 0°.

[0055] Figure 9 The mean FMC (Fragrant Mid-Capacity) values ​​for different vegetation types in each month of the year from 2010 to 2021 were statistically analyzed for different temperature zones. The graph is divided into three parts, corresponding to the tropical, northern temperate, and southern temperate zones, respectively. Due to missing data for other zones, this study does not analyze them. The different colored lines in the graph represent four vegetation types: grassland, broadleaf forest, coniferous forest, and shrubland.

[0056] In tropical regions ( Figure 9 In the North Temperate Zone (A), grasslands exhibit the least fluctuation in free-floating molecular weight (FMC), typically between 45% and 60%. Broadleaf forests show relatively stable FMC, ranging from 75% to 100%, with significant seasonal variations. Coniferous forests and shrubs show greater FMC fluctuations, with shrub FMC ranging from 60% to 130%, exhibiting very pronounced seasonal variations, particularly peaking at the beginning and end of each year. Figure 9 In species B, grasslands exhibit relatively stable FMC (free mulch content), ranging from 50% to 70%. Broadleaf and coniferous forests show greater FMC fluctuations, especially coniferous forests, where seasonal FMC changes are very pronounced, ranging from 100% to 150%. Shrubs show significant FMC fluctuations, ranging from 70% to 200%, indicating significant annual humidity variations; in the southern temperate zone (… Figure 9 C) Grassland FMC fluctuates less, typically between 50% and 80%. Broadleaf and coniferous forests exhibit greater FMC fluctuations, with coniferous forest FMC fluctuating between 80% and 140%.

[0057] The seasonal fluctuations in FMC of shrubs are also quite significant, typically ranging from 80% to 130%.

[0058] All vegetation types exhibit significant seasonal fluctuations in water content (FMC), particularly in tropical and temperate regions. Shrub and coniferous forests show greater fluctuations in FMC, especially in temperate regions, where the seasonal variation is more pronounced than in grasslands and broadleaf forests. Compared to temperate regions, tropical regions show less FMC fluctuation, but shrub and coniferous forests still exhibit clear seasonal trends.

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] Table 6 Vegetation Index Information

[0065] Note: R λ The reflectance at a wavelength of λ nm; λ 1. λ 2 and λ 3 refers to the wavelengths at three locations, which in this application are the corresponding band values ​​obtained after processing by the MODSI system for bands 2, 5, and 7. r and w These are two calculated spectra, which in this application represent the moisture absorption coefficient and the simulated reflectance used in the model simulation process, respectively; λ represents the band position; and L is the number of spectral bands within the spectral range.

Claims

1. A method for retrieving the moisture content of combustibles based on a mechanistic model and artificial neural networks, mainly comprising the following steps: S1: Construct a simulated dataset of canopy reflectance based on the IGBP classification system; S11: The reflectance of grassland vegetation canopy was simulated using the PROSAIL model, with a simulation band of 400-2500 nm and a spectral resolution of 1 nm. S12: The reflectance of non-uniform vegetation canopy was simulated using a 4-Scale geometric optics model. The input parameters of the 4-Scale geometric optics model included the leaf reflectance and transmittance obtained from the PROSPECT model. The simulation band was 400-2500 nm and the spectral resolution was 1 nm. S13: Using the functional relationship between the equivalent water thickness, dry matter content, chlorophyll content, and leaf area index (LAI) of leaves as constraints to eliminate unreasonable data, the range of variation of the equivalent water thickness, dry matter content, and chlorophyll content of leaves is limited as follows: In formulas (2) and (3), V 𝑚𝑖𝑛 (LAI), V max (LAI) represents the minimum and maximum values ​​of one of the following parameters of leaf equivalent water thickness, dry matter content, and chlorophyll content, respectively, as LAI changes. 𝑚𝑖𝑛 V 𝑚𝑎𝑥 These represent the minimum and maximum values ​​of the parameter when LAI reaches its minimum value, respectively, and V. min (LAI 𝑚𝑎𝑥 V 𝑚𝑎𝑥 (LAI 𝑚𝑎𝑥 These represent the maximum and minimum values ​​of the parameter when LAI reaches its maximum value; S14: Based on Gaussian white noise, add additive and multiplicative uncertainties, including band-correlated and band-independent uncertainties, to the simulated reflectance data: In formula (4), where and These represent the reflectivity values ​​of the λ-band simulated by the radiative transfer model and the reflectivity values ​​after adding noise, respectively. Represents a normal distribution. It is a relative uncertainty that applies to all wavebands. This refers to the relative uncertainty applied to the λ-band. It is an absolute uncertainty that applies to all wavebands. This refers to the absolute uncertainty applied to the λ-band. = = 0.01, = = 4%; S15: The processed simulated reflectance data is converted according to the spectral response function of MODIS to obtain simulated canopy reflectance data for 7 bands in the corresponding MODIS wavelength range. S16: Vegetation indices related to equivalent water thickness and dry matter content are calculated based on the simulated canopy reflectance of seven bands in MODIS. S2: Construct an FMC inversion model based on the IGBP classification system; S21: The simulated datasets for the four vegetation types are divided into two parts. One part is used for neural network training, and the other part is used to test the performance of the mechanistic model. The simulated datasets include simulated canopy reflectance, vegetation index, and leaf area index in bands 1 to 7. The mechanistic model is expressed as follows: In formula (1), FMC is the combustible water content of the vegetation canopy, EWT is the equivalent water thickness of the vegetation canopy, and DMC is the dry matter content of the vegetation canopy. S22: Train the neural network until the output of the mechanism model meets the requirements; S3: Achieve FMC inversion through a global vegetation status monitoring system based on the IGBP classification system.

2. The method for retrieving the moisture content of combustibles based on a mechanistic model and artificial neural network according to claim 1, characterized in that: Step S22 includes S221 input layer optimization: determining the optimal input parameters by comparing the importance of each parameter to determine the optimal band combination, given that the output layer, hidden layer, and other model parameters are fixed; parameter selection uses a sequential forward search algorithm to select the subset that contributes most to the performance of the mechanistic model from the given feature set. During parameter selection, the maximum coefficient of determination R of the two parameters EWT and DMC in the mechanistic model results is selected. 2 The optimal band combination with the minimum normalized root mean square error (nRMSE) is used as the basis for the next parameter increase, until the parameter selection meets the requirements.

3. The method for retrieving the moisture content of combustibles based on a mechanistic model and artificial neural network according to claim 2, characterized in that: The S22 step includes S222 output layer optimization: with the input layer, hidden layer and other parameters of the mechanism model fixed, the performance of the mechanism model is compared when the five output variables EWT, DMC, Car, Cab, and Ant, the two co-output variables EWT and DMC, and the single variables EWT and DMC are used as outputs.

4. The method for retrieving the moisture content of combustibles based on a mechanistic model and artificial neural network according to claim 3, characterized in that: Step S22 includes hidden layer optimization in step S223: tuning the hidden layer parameters by continuously adjusting the activation function, learning rate, and optimizer parameters of the mechanism model to obtain the optimal mechanism model parameter information. The number of neurons in the three hidden layers are 64, 128, and 64, respectively. The loss function is mean squared error. The activation function of each hidden layer is the Tanh function, the activation function of the output layer is the Linear function, the optimization function of the output layer is the Adam function, and the learning rate of the output layer is 0.

005.

5. The method for retrieving the moisture content of combustibles based on a mechanistic model and an artificial neural network according to claim 2, characterized in that: The parameters in the S221 input layer optimization step include vegetation index parameters, which include normalized differential infrared index, normalized vegetation index, enhanced vegetation index, water stress index, normalized differential water index, simple ratio water index, shortwave infrared water stress index, normalized spectral angle index, and similarity water index.

6. The method for inverting the moisture content of combustibles based on a mechanistic model and artificial neural network according to any one of claims 1 to 5, characterized in that: It also includes step S4. S4: Verify the inversion results and evaluate the performance of the FMC inversion model.

7. The method for inverting the moisture content of combustibles based on a mechanistic model and artificial neural network according to claim 1, characterized in that: The equivalent water thickness, dry matter content, and chlorophyll content of the leaves in step S13 are limited to the corresponding data ranges in Table 5.

8. The method for retrieving the moisture content of combustibles based on a mechanistic model and artificial neural network according to claim 1, characterized in that: The global vegetation status monitoring system based on the IGBP classification system uses Google Earth Engine.

9. The application of the combustible material moisture content inversion method based on mechanistic model and artificial neural network according to any one of claims 1 to 8, characterized in that: Acquisition and representation of the spatial distribution characteristics of global FMC.

10. The application of the combustible material moisture content inversion method based on mechanistic model and artificial neural network according to claim 9, characterized in that: Acquisition and representation of the temporal distribution characteristics of global FMC.

Images