A combustible moisture content inversion method based on multi-layer soil moisture time sequence characteristics
Patent Information
- Application Number
- CN202611221778.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-12
- Publication Date
- 2026-10-16
AI Technical Summary
若仅采用单时相或统一时间窗口提取特征，难以刻画其真实时序规律，从而限制反演精度
[0039] The beneficial effects of this invention are as follows: By introducing a multi-layer soil water time series feature construction method, it can systematically characterize the historical evolution of soil water changes at different depths and their cumulative impact on vegetation water status, thereby significantly enhancing the model's ability to express the dynamic change mechanism of vegetation water; by constructing a dynamic time window optimization mechanism for different vegetation types and soil depths, it achieves adaptive selection of time scales, effectively solving the problems of time series information distortion and lag response mismatch caused by fixed time windows in traditional methods, and improving the targeting and rationality of feature extraction under different ecosystem conditions; by introducing a model-based... The feature contribution analysis mechanism of the model interpretation performs differential screening of high-dimensional time series features, effectively eliminating redundant and low-contribution features, reducing feature space complexity, and improving model training efficiency and prediction stability. At the same time, a differentiated machine learning model system is constructed for different vegetation types, so that forests, shrubs and grasslands can adopt the model structure most suitable for their data distribution characteristics, thereby further improving the model's adaptability and generalization performance in complex ecological environments. In summary, this invention can achieve high-precision, spatiotemporally continuous remote sensing inversion of vegetation canopy combustible water content, and has strong engineering application value and promotion significance.
Smart Images
Figure CN122926337A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of remote sensing inversion, machine learning, and time series data analysis. It relates to a remote sensing inversion method for vegetation canopy combustible moisture content (FMC) based on multi-layer soil moisture time series characteristics, which can be applied to forest fire risk assessment, fire risk early warning, and dynamic monitoring of vegetation moisture. Background Technology
[0002] Fuel moisture content (FMC) of vegetation canopy is a key physiological parameter characterizing vegetation moisture status and flammability, and is also an important input variable for simulating forest fire occurrence, spread, and fire behavior. FMC directly reflects the degree of vegetation response to drought stress, and its changes are influenced by multiple factors, including vegetation type, soil moisture, and meteorological conditions, exhibiting significant temporal dynamic characteristics. Therefore, constructing a high-precision, spatiotemporally continuous FMC remote sensing inversion method has significant theoretical and practical value for forest fire risk prediction, fire risk assessment, and ecological environment monitoring.
[0003] Currently, FMC remote sensing inversion methods mainly include physical model methods based on radiative transfer mechanisms and data-driven methods based on statistical learning or machine learning. Physical models establish the radiative transfer relationship between remote sensing observations and vegetation physiological parameters, offering good physical interpretability. However, they suffer from problems such as numerous parameters, complex structures, high computational costs, and strong dependence on input data quality, leading to low efficiency and incomplete results in large-scale applications. In contrast, machine learning methods can utilize multi-source remote sensing data and environmental variables to construct complex nonlinear mapping relationships, offering advantages such as high computational efficiency, strong adaptability, and ease of generalization, making them an important development direction for FMC inversion.
[0004] Existing machine learning methods typically use soil moisture, vegetation indices, and meteorological variables as static input features for direct modeling, neglecting the lagged response characteristics of vegetation to changes in soil moisture. In reality, vegetation water content depends not only on the soil moisture state at the time of observation but also on the soil moisture accumulation process over a certain period. Furthermore, different vegetation types exhibit significant differences in response time to soil moisture in different soil layers due to variations in root depth, water intake methods, and ecological adaptability; forests, shrubs, and grasslands demonstrate different temporal scale characteristics. Using only a single temporal phase or a uniform time window to extract features fails to capture the true temporal patterns, thus limiting the accuracy of inversion.
[0005] Furthermore, existing methods often employ fixed time windows (such as 30, 90, or 180 days) in constructing time series features, failing to consider differences in vegetation type and soil layer, which can easily lead to missing or redundant information. Simultaneously, as feature dimensions increase, the correlation between variables strengthens, and directly using these features for model training can easily result in the accumulation of redundant features, increased model complexity, and decreased generalization ability. Therefore, there is an urgent need for a method that integrates multi-layer soil water time series features, employs a dynamic time window optimization mechanism tailored to different vegetation types and soil layers, and combines feature importance analysis to achieve differentiated feature selection, thereby improving the accuracy, stability, and regional applicability of FMC remote sensing inversion. Summary of the Invention
[0006] This invention proposes a remote sensing inversion method for vegetation canopy combustible moisture content (FMC) that integrates multi-layer soil water time-series feature construction, dynamic time window optimization, and a SHAP-based feature selection mechanism. This method utilizes four layers of ERA5-Land soil moisture data, vegetation type data, and ground-measured FMC data to construct multi-layer soil water time-series features and optimizes the time window for different vegetation types and soil layers. By combining machine learning models such as TabPFN and TabNet, a nonlinear mapping relationship between soil moisture time-series changes and FMC is established. The model is optimized through feature importance analysis, achieving high-precision, spatiotemporally continuous remote sensing inversion of FMC for forests, shrubs, and grasslands.
[0007] To achieve the above objectives, the present invention provides the following technical solution:
[0008] A method for inverting combustible moisture content based on multi-layer soil moisture temporal characteristics includes the following steps:
[0009] Step 1: Obtain soil moisture data, vegetation type data, and measured data of canopy combustible water content (FMC) of the four layers of ERA5-Land. Perform spatial matching, temporal matching, and quality control processing on the multi-source data to construct a standardized training dataset.
[0010] Step 2: Based on the soil moisture time series at different soil depths, extract trend features, statistical features, and dynamic time series features to construct a multi-layer soil water time series feature vector;
[0011] Step 3: Set candidate time windows for different vegetation types and soil depths, and determine the optimal combination of time windows for each vegetation type and soil layer through model performance evaluation;
[0012] Step 4: Using the SHAP feature contribution analysis method, the importance of the feature vectors of multi-layer soil water time series is evaluated, low-contribution and redundant features are screened out, and the optimal feature subsets corresponding to different vegetation types are constructed.
[0013] Step 5: Based on the optimal time window and the optimal feature subset, construct machine learning models for forest, shrub and grassland vegetation types respectively to achieve prediction and spatial continuous inversion of the moisture content of vegetation canopy combustibles.
[0014] Furthermore, step one is described in detail as follows:
[0015] Measured data of combustible water content in the vegetation canopy were obtained as labels for supervised learning, and the corresponding sampling time and geographical location information were recorded. Simultaneously, soil moisture data from four layers of ERA5-Land and IGBP land cover type data were acquired to determine vegetation type. Consistency processing was performed on the multi-source data. First, the ERA5-Land and IGBP data were resampled to a uniform spatial resolution. Second, time matching was performed based on the FMC sampling date to extract soil moisture time series data prior to the corresponding date. Finally, outliers and missing values were removed to obtain a standardized sample set.
[0016] Furthermore, the trend feature uses the least squares method to fit the time series change trend of volumetric moisture content, and reflects the overall change benchmark of the time series based on the linear fitting intercept; the statistical feature integrates the standard deviation, median, maximum value, and minimum value of volumetric moisture content; the complex time series feature introduces the number of sign changes, absolute energy, and third-order delay correlation features to characterize the dynamic change law of volumetric moisture content.
[0017] The above three features were extracted again for the four soil depth layers of ERA5-Land to form multidimensional time series feature vectors for each soil layer. The four layers of features were then spliced and fused to construct multi-layer soil water time series feature vectors.
[0018] Furthermore, step three is detailed as follows:
[0019] The time window optimization process was performed independently for each of the three vegetation types (forest, shrubland, and grassland) and the four soil depth layers of ERA5-Land:
[0020] Based on the FMC observation date, multiple candidate historical time window sets were set for the ERA5-Land four-layer soil moisture data. Under any vegetation type and soil layer condition, the trend features, statistical features and dynamic time series features were extracted in each candidate window. With the optimization objectives of maximizing model accuracy and minimizing error, a vegetation type-adapted machine learning model was used for training and validation. The candidate windows were comprehensively compared to determine the optimal time window.
[0021] Furthermore, step four is detailed as follows:
[0022] For three vegetation types—forest, shrub, and grassland—feature contribution analysis was performed using a vegetation type-adapted machine learning model. The contribution value of each input feature in the multi-layer soil water time series feature vector was calculated, and the optimal feature subset was constructed based on feature contribution ranking and redundancy analysis.
[0023] Furthermore, step five is detailed as follows:
[0024] The optimal time window and the optimal feature subset are input into the vegetation type-adapted machine learning model to establish a nonlinear mapping relationship between the combustible water content of the vegetation canopy and the input features. After the model is trained, the prediction results of different vegetation type models are uniformly standardized and fused according to spatial distribution to generate regional scale FMC continuous distribution results covering different vegetation types.
[0025] Furthermore, the ERA5-Land four-layer soil moisture data includes daily volumetric water content data for four soil depth layers: 0–7 cm, 7–28 cm, 28–100 cm, and 100–289 cm.
[0026] Furthermore, the specific machine learning model for vegetation type adaptation is as follows: the TabPFN model is used for forest and shrub types, and the TabNet model is used for grassland types.
[0027] Furthermore, the volumetric water content of the l-th soil depth layer at time t is expressed as SM. l (t), where l represents the soil depth layer number and t represents the time series index; the trend feature is fitted using the least squares method to fit the time series change trend, and its slope is defined as:
[0028]
[0029] Where, n is the length of the time window. The mean of the time series index. Slope represents the mean soil moisture, and is used to describe the rate of change of soil moisture over time.
[0030] Simultaneously, the linear fit intercept is calculated to reflect the overall change benchmark of the time series:
[0031]
[0032] The complex timing features include the number of symbol changes. Absolute energy The third-order delay correlation feature C3 is calculated as follows:
[0033]
[0034]
[0035]
[0036] Where I(⋅) is an indicator function used to count the number of times the direction of sequence change changes, thereby reflecting the frequency of soil moisture fluctuations.
[0037] Furthermore, the candidate window is set within a range of 10–270 days.
[0038] After model training, the prediction results of models for different vegetation types are standardized and fused according to spatial distribution to generate continuous regional-scale FMC distribution results covering different vegetation types such as forests, shrubs, and grasslands. By modeling by vegetation type and combining dynamic time window optimization and feature selection mechanisms, the model can adopt optimal input features and prediction models for different ecosystems, effectively improving the accuracy, stability, and spatial generalization ability of remote sensing inversion of vegetation canopy combustible water content.
[0039] The beneficial effects of this invention are as follows: By introducing a multi-layer soil water time series feature construction method, it can systematically characterize the historical evolution of soil water changes at different depths and their cumulative impact on vegetation water status, thereby significantly enhancing the model's ability to express the dynamic change mechanism of vegetation water; by constructing a dynamic time window optimization mechanism for different vegetation types and soil depths, it achieves adaptive selection of time scales, effectively solving the problems of time series information distortion and lag response mismatch caused by fixed time windows in traditional methods, and improving the targeting and rationality of feature extraction under different ecosystem conditions; by introducing a model-based... The feature contribution analysis mechanism of the model interpretation performs differential screening of high-dimensional time series features, effectively eliminating redundant and low-contribution features, reducing feature space complexity, and improving model training efficiency and prediction stability. At the same time, a differentiated machine learning model system is constructed for different vegetation types, so that forests, shrubs and grasslands can adopt the model structure most suitable for their data distribution characteristics, thereby further improving the model's adaptability and generalization performance in complex ecological environments. In summary, this invention can achieve high-precision, spatiotemporally continuous remote sensing inversion of vegetation canopy combustible water content, and has strong engineering application value and promotion significance. Attached Figure Description
[0040] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:
[0041] Figure 1 This is a schematic diagram of the overall method flow of the present invention;
[0042] Figure 2This is a schematic diagram of the time window exploration results under different vegetation types and different soil depths in this invention, wherein: (a) forest, (b) shrubs, and (c) grassland;
[0043] Figure 3 This is a schematic diagram of the feature importance analysis results based on the SHAP method of the present invention, wherein: (a) forest, (b) shrubs, and (c) grassland;
[0044] Figure 4 The model inversion accuracy results are for different vegetation types, including: (a) forest, (b) shrubs, and (c) grassland. Detailed Implementation
[0046] The invention will be further described below with reference to the accompanying drawings and real-world examples.
[0047] This embodiment provides a method for inverting the moisture content of combustibles based on multi-layer soil moisture temporal characteristics, such as... Figure 1 As shown, it includes the following steps:
[0048] Step 1: Obtain soil moisture data, vegetation type data, and measured data of canopy combustible water content (FMC) of the four layers of ERA5-Land. Perform spatial matching, temporal matching, and quality control processing on the multi-source data to construct a standardized training dataset.
[0049] Specifically, measured data of vegetation canopy combustible moisture content (FMC) were acquired as supervised learning labels, and the corresponding sampling time and geographical location information were recorded. Simultaneously, ERA5-Land reanalysis soil moisture data were acquired, including daily volumetric water content data for four soil depths: 0–7 cm, 7–28 cm, 28–100 cm, and 100–289 cm. IGBP land cover type data were also acquired to determine vegetation type (forest, shrub, and grassland). Consistency processing was performed on the multi-source data. First, the ERA5-Land and IGBP data were resampled to a spatial resolution of 10 km. Second, time matching was performed based on the FMC sampling date to extract soil moisture time series data prior to that date. Finally, outliers and missing values were removed to obtain a standardized sample set for subsequent model training and analysis.
[0050] Step 2: Based on the soil moisture time series at different soil depths, extract trend features, statistical features, and dynamic time series features to construct a multi-layer soil water time series feature vector.
[0051] Specifically, based on the four-layer soil moisture data of ERA5-Land, a multi-layer time series feature model capable of characterizing the dynamic response of vegetation to soil moisture was constructed. The volumetric water content of the l-th soil depth layer at time t is denoted as SM.l (t), where l represents the soil depth layer number and t represents the time series index. Continuous soil moisture sequences for each soil layer are extracted within a preset time window based on the FMC observation date, and corresponding time series features are constructed. For this soil moisture sequence, features are extracted from three levels: trend features, statistical features, and dynamic time series features, forming a multi-layered soil water time series feature vector.
[0052] The trend features are fitted using the least squares method to determine the trend of the time series, and their slope is defined as follows:
[0053]
[0054] Where l represents the soil depth layer number, t is the time series position, n is the time window length, and SM l (t) represents the soil moisture content of the l-th layer at time t. Slope is used to describe the rate of change of soil moisture over time.
[0055] Simultaneously, the linear fit intercept is calculated to reflect the overall change benchmark of the time series:
[0056]
[0057] The statistical features characterize the overall distribution of the soil moisture sequence, and the standard deviation of volumetric water content is calculated respectively. , median Maximum value Minimum value :
[0058]
[0059]
[0060]
[0061]
[0062] The aforementioned complex temporal features further characterize the dynamic changes in soil moisture sequences, introducing the sign change count. Absolute Energy The time series features, including the third-order delay correlation feature (C3), are calculated as follows:
[0063]
[0064]
[0065]
[0066] Where I(⋅) is an indicator function used to count the number of times the direction of sequence change changes, thereby reflecting the frequency of soil moisture fluctuations.
[0067] Based on this, the above feature extraction process was repeated for each of the four soil depth layers in ERA5-Land, forming multi-dimensional time-series feature vectors for each soil layer. The four feature layers were then concatenated and fused to construct a unified feature representation that comprehensively reflects the dynamic changes in soil moisture across different soil layers. This multi-layered soil water time-series feature vector can enhance the characterization of the lagged response and cumulative effect of vegetation moisture changes while preserving historical soil moisture evolution information. This provides an input basis for subsequent dynamic time window optimization, feature selection, and high-precision inversion of vegetation canopy combustible moisture content.
[0068] Step 3: Set candidate time windows for different vegetation types and soil depths, and determine the optimal combination of time windows for each vegetation type and soil layer through model performance evaluation.
[0069] Specifically, based on the FMC observation date, multiple candidate historical time window sets W={w1,w2,…,w...} are set for the four-layer soil moisture series of ERA5-Land. n}, where w i Different window lengths are represented, and the window is expanded step by step according to a preset step size. In this embodiment, the candidate window range is set to 10–270 days. Under any vegetation type and soil layer condition, the time series features described in step two are extracted within each candidate window, and a corresponding training dataset is constructed. The machine learning model is adapted using the same type of vegetation as the final FMC inversion model in step five for training and validation. The model performance indicators, the coefficient of determination R² and the root mean square error RMSE, are calculated to evaluate the inversion effect under different windows.
[0070]
[0071]
[0072] in, This indicates the measured FMC. This represents the model's predicted FMC value, and N represents the number of samples.
[0073] In this embodiment, corresponding machine learning models are used for different vegetation types. Forest and shrub types use the TabPFN (Tabular Prior-data Fitted Network) model, while grassland types use the TabNet (Attentive Interpretable Tabular Learning Network) model. Steps three and five use the same type of vegetation adaptation model. The difference is that step three uses the feature set corresponding to the candidate time windows to evaluate model performance and determine the optimal time window combination; step five uses the optimal time window determined in step three and the optimal feature set selected in step four to retrain the model for the final FMC inversion.
[0074] With the optimization objectives of maximizing model accuracy (maximum R²) and minimizing error (minimum RMSE), candidate windows are comprehensively compared to determine the optimal time window w*. When the performance of multiple windows is similar, the shorter and more stable window is selected first to reduce redundant historical information and improve computational efficiency.
[0075] Furthermore, this invention independently performs time window optimization processes for three vegetation types (forest, shrub, and grassland) and four soil depth layers (ERA5-Land), constructing a correspondence between "vegetation type - soil layer - time window" to achieve adaptive matching of soil moisture response time scales for different ecological types.
[0076]
[0077] Where w∗ represents the optimal combination of time windows determined for a specific vegetation type and soil depth layer; This indicates the vegetation type (v), soil depth (l), and candidate time window (w). i Under the given conditions, the coefficient of determination between the prediction results of the machine learning model and the measured FMC; This indicates the search for the coefficient of determination R. 2 The largest combination of time windows.
[0078] like Figure 2 As shown, different vegetation types exhibit significantly different optimal time response scales under different soil conditions. Forest vegetation, due to its deeper root system, has a longer cumulative response to deep soil moisture over a longer time scale. Shrub vegetation is more sensitive to changes in soil moisture over a medium time scale, while grassland vegetation responds more significantly to changes in short-term soil moisture.
[0079] Table 1: Range of time steps for different vegetation types at different soil depths
[0080] forest 120 160 190 220 shrub 90 110 160 170 grassland 90 110 130 170
[0081] Step 4: Utilize the SHAP (SHapley Additive exPlanations) feature contribution analysis method to evaluate the importance of the multi-layered soil water time series feature vectors constructed in Step 2. Low-contribution and redundant features are removed, and optimal feature subsets corresponding to different vegetation types are constructed.
[0082] To reduce redundancy among multi-layered soil water time series features, importance analysis and screening are performed on the features based on the determination of the optimal time window, thereby reducing feature redundancy and improving the model's generalization ability. For example... Figure 3 As shown, this invention analyzes the feature contributions of three vegetation types—forest, shrub, and grassland—using a vegetation type-adapted machine learning model identical to the final FMC inversion model. The TabPFN model is used for forest and shrub types, while the TabNet model is used for grassland. The contribution value of each input feature in the multi-layer soil water time series feature vector constructed in step two is calculated using the aforementioned model interpretation method, and then sorted from highest to lowest contribution value to form a feature importance sequence. Based on the feature importance ranking results and considering changes in model prediction performance, features with low contribution or significant redundancy are gradually eliminated, retaining features that significantly contribute to model prediction, thus constructing an optimal feature subset. Since forest, shrub, and grassland differ in soil water use patterns and response mechanisms, this invention conducts feature contribution analysis separately for each of the three vegetation types and, combined with the importance results of the four-layer soil water time series features, establishes corresponding differentiated feature combinations, enabling different vegetation types to use the most suitable input features for model training.
[0083] Step 5: Based on the optimal time window and the optimal feature subset, construct machine learning models for forest, shrub and grassland vegetation types respectively to achieve prediction and spatial continuous inversion of the moisture content of vegetation canopy combustibles.
[0084] Specifically, based on the differences in ecological structure and sample distribution among forests, shrubs, and grasslands, corresponding machine learning models were established. Forests and shrubs were modeled using the TabPFN model, while grasslands were modeled using the TabNet model, to fully leverage the learning capabilities of different models for various vegetation features. Feature vectors were input into the corresponding vegetation type's machine learning model to establish a non-linear mapping relationship between the canopy combustible water content and the input features.
[0085] After model training, the prediction results of models for different vegetation types are standardized and fused according to spatial distribution to generate continuous regional-scale FMC distribution results covering different vegetation types such as forests, shrubs, and grasslands. By modeling by vegetation type and combining dynamic time window optimization and feature selection mechanisms, the model can adopt optimal input features and prediction models for different ecosystems, effectively improving the accuracy, stability, and spatial generalization ability of remote sensing inversion of vegetation canopy combustible water content.
[0086] like Figure 4 As shown, the accuracy of the models for different vegetation types was verified in this invention. The results are as follows: Forest type: R² = 0.72, RMSE = 17.98, sample size 5894. Shrub type: R² = 0.72, RMSE = 12.50, sample size 4318. Grassland type: R² = 0.51, RMSE = 23.43, sample size 44935. The results indicate that the method of this invention exhibits good inversion accuracy under different ecosystem conditions. Among them, forest and shrub types have better fitting effects, while grassland type maintains a moderate level of accuracy due to its strong spatial heterogeneity. However, the overall models all have strong stability and generalization ability.
[0087] Based on the trained model, pixel-by-pixel predictions were made on global ERA5-Land soil moisture data and corresponding vegetation types to obtain the spatial distribution results of vegetation canopy combustible moisture content at a global scale. The results show that the present invention can effectively characterize the differences in vegetation moisture in different regions and has good spatial continuity and ecological rationality.
[0088] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.
Claims
1. A method for inverting the moisture content of combustibles based on multi-layer soil moisture temporal characteristics, characterized in that, Includes the following steps: Step 1: Obtain soil moisture data, vegetation type data, and measured data of canopy combustible water content (FMC) of the four layers of ERA5-Land. Perform spatial matching, temporal matching, and quality control processing on the multi-source data to construct a standardized training dataset. Step 2: Based on the soil moisture time series at different soil depths, extract trend features, statistical features, and dynamic time series features to construct a multi-layer soil water time series feature vector; Step 3: Set candidate time windows for different vegetation types and soil depths, and determine the optimal combination of time windows for each vegetation type and soil layer through model performance evaluation; Step 4: Using the SHAP feature contribution analysis method, the importance of the feature vectors of multi-layer soil water time series is evaluated, low-contribution and redundant features are screened out, and the optimal feature subsets corresponding to different vegetation types are constructed. Step 5: Based on the optimal time window and the optimal feature subset, construct machine learning models for forest, shrub and grassland vegetation types respectively to achieve prediction and spatial continuous inversion of the moisture content of vegetation canopy combustibles.
2. The method for inverting combustible material moisture content based on multi-layer soil moisture temporal characteristics according to claim 1, characterized in that, Step one is described in detail as follows: Measured data of combustible water content in the vegetation canopy were obtained as labels for supervised learning, and the corresponding sampling time and geographical location information were recorded. Simultaneously, soil moisture data from four layers of ERA5-Land and IGBP land cover type data were acquired to determine vegetation type. Consistency processing was performed on the multi-source data. First, the ERA5-Land and IGBP data were resampled to a uniform spatial resolution. Second, time matching was performed based on the FMC sampling date to extract soil moisture time series data prior to the corresponding date. Finally, outliers and missing values were removed to obtain a standardized sample set.
3. The method for inverting combustible material moisture content based on multi-layer soil moisture temporal characteristics according to claim 1, characterized in that, The trend features use the least squares method to fit the time series change trend of volumetric moisture content, and reflect the overall change benchmark of the time series based on the linear fitting intercept; the statistical features integrate the standard deviation, median, maximum value and minimum value of volumetric moisture content; the complex time series features introduce the number of sign changes, absolute energy and third-order delay correlation features to characterize the dynamic change law of volumetric moisture content. The above three features were extracted again for the four soil depth layers of ERA5-Land to form multidimensional time series feature vectors for each soil layer. The four layers of features were then spliced and fused to construct multi-layer soil water time series feature vectors.
4. The method for inverting combustible material moisture content based on multi-layer soil moisture temporal characteristics according to claim 1, characterized in that, Step three is described in detail below: The time window optimization process was performed independently for each of the three vegetation types (forest, shrubland, and grassland) and the four soil depth layers of ERA5-Land: Based on the FMC observation date, multiple candidate historical time window sets were set for the ERA5-Land four-layer soil moisture data. Under any vegetation type and soil layer condition, the trend features, statistical features and dynamic time series features were extracted in each candidate window. With the optimization objectives of maximizing model accuracy and minimizing error, a vegetation type-adapted machine learning model was used for training and validation. The candidate windows were comprehensively compared to determine the optimal time window.
5. The method for inverting combustible material moisture content based on multi-layer soil moisture temporal characteristics according to claim 1, characterized in that, Step four is described in detail below: For three vegetation types—forest, shrub, and grassland—feature contribution analysis was performed using a vegetation type-adapted machine learning model. The contribution value of each input feature in the multi-layer soil water time series feature vector was calculated, and the optimal feature subset was constructed based on feature contribution ranking and redundancy analysis.
6. The method for inverting combustible material moisture content based on multi-layer soil moisture temporal characteristics according to claim 1, characterized in that, Step five is described in detail below: The optimal time window and the optimal feature subset are input into the vegetation type-adapted machine learning model to establish a nonlinear mapping relationship between the combustible water content of the vegetation canopy and the input features. After the model is trained, the prediction results of different vegetation type models are uniformly standardized and fused according to spatial distribution to generate regional scale FMC continuous distribution results covering different vegetation types.
7. The method for inverting combustible material moisture content based on multi-layer soil moisture temporal characteristics according to claim 1, characterized in that, The ERA5-Land four-layer soil moisture data includes daily volumetric water content data for four soil depth layers: 0–7 cm, 7–28 cm, 28–100 cm, and 100–289 cm.
8. The method for inverting combustible material moisture content based on multi-layer soil moisture temporal characteristics according to claim 1, characterized in that, The specific machine learning models adapted to the vegetation types are as follows: the TabPFN model is used for forest and shrub types, and the TabNet model is used for grassland types.
9. The method for inverting combustible material moisture content based on multi-layer soil moisture temporal characteristics according to claim 3, characterized in that, The volumetric water content of the l-th soil depth layer at time t is denoted as SM. l (t), where l represents the soil depth layer number and t represents the time series index; the trend feature is fitted using the least squares method to fit the time series change trend, and its slope is defined as: Where, n is the length of the time window. The mean of the time series index. Slope represents the mean soil moisture, and is used to describe the rate of change of soil moisture over time. Simultaneously, the linear fit intercept is calculated to reflect the overall change benchmark of the time series: The complex timing features include the number of symbol changes. Absolute energy The third-order delay correlation feature C3 is calculated as follows: Where I(⋅) is an indicator function used to count the number of times the direction of sequence change changes, thereby reflecting the frequency of soil moisture fluctuations.
10. The method for inverting combustible material moisture content based on multi-layer soil moisture temporal characteristics according to claim 4, characterized in that, The candidate window is set within a range of 10–270 days.