Method for estimating the load of dead combustible material on the surface of subtropical forests using multi-source remote sensing and machine learning
By combining multi-source remote sensing data and machine learning models, the problem of canopy shading in subtropical forests has been solved, enabling accurate estimation of surface dead combustible load and improving the accuracy of fire risk assessment and management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JIANGXI NORMAL UNIV
- Filing Date
- 2025-12-24
- Publication Date
- 2026-07-03
AI Technical Summary
In subtropical forests, dense canopy cover makes it difficult for optical remote sensing to accurately capture signals of dead combustibles on the ground. Existing models are inadequate in handling complex nonlinear datasets, affecting forest fire risk assessment and combustible management.
By combining multi-source remote sensing data (Landsat 8 OLI, Sentinel-1 C-band SAR, ALOS-2 PALSAR-2 L-band SAR) with machine learning models (XGBoost, LightGBM), feature extraction and variable selection are performed to construct a combustible load prediction model and conduct regional inversion.
It provides robust and spatially defined SDFL estimates, improving the accuracy of forest fire risk assessment and the precision of combustible material management. The machine learning model outperforms the linear regression model, especially LightGBM, which performs best across multiple SDFL components.
Smart Images
Figure CN121765685B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of forest fire monitoring technology, and in particular to a method for estimating the surface dead combustible load of subtropical forests using multi-source remote sensing and machine learning. Background Technology
[0002] Accurate estimation of surface dead fuel load (SDFL) is crucial for effective forest fire management. In remote sensing-based estimations, researchers use optical remote sensing to indirectly estimate litter decomposition rates and surface fuel dynamics through ecological processes, then combine litter accumulation with iterative decomposition rate estimations to quantify regional SDFL. Other studies have used optical remote sensing to directly estimate SDFL, but these studies have primarily focused on areas with simple vegetation structures, including savanna and desert-steppe transition zones. However, in subtropical forests, dense canopy cover (vegetation cover typically exceeding 0.7) leads to significant canopy blockage, posing a challenge to capturing signals from forest floor dead fuel using optical remote sensing alone. To address this limitation, studies in areas with complex forest structures increasingly employ estimation methods combining UAV-mounted lidar with optical remote sensing. While lidar can accurately capture the relationship between vegetation structure and SDFL, airborne lidar applications are limited by limited coverage, hindering large-scale prediction and inversion. C-band synthetic aperture radar can penetrate moderately dense canopies, but is still limited by canopy top scattering in highly enclosed areas. In contrast, the strong penetration capability of L-band SAR compensates for this limitation, enabling direct acquisition of surface dead combustible load signals. Therefore, multi-source data fusion combining optical remote sensing (reflecting canopy and land cover characteristics) with C / L-band SAR (reflecting understory structure) is key to overcoming the "canopy shading" bottleneck in subtropical forests. Furthermore, significant differences exist in productivity, organ shedding patterns, and mortality cycles among different vegetation types, directly affecting the quantity and composition of SDFL inputs. Simultaneously, topography (elevation, slope, aspect) synergistically influences SDFL accumulation and transport.
[0003] Currently, typical models used for SDFL prediction include Multiple Linear Regression (MLR) and Random Forest (RF). However, the performance of Extreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM) in SDFL prediction has not been fully explored. Each of these models has different advantages and limitations. For example, while MLR models can intuitively express the relationships between variables, their ability to model nonlinear patterns is limited. In contrast, Random Forest has strong nonlinear fitting and generalization capabilities, but it is prone to overfitting when dealing with noisy datasets. Therefore, when dealing with complex nonlinear datasets, choosing an appropriate model and optimizing its parameters is crucial. Summary of the Invention
[0004] The purpose of this invention is to provide a method for estimating the surface dead combustible load of subtropical forests using multi-source remote sensing and machine learning, combining multi-source remote sensing with machine learning to provide methodological support for fire risk assessment and precise combustible management in subtropical forests.
[0005] To achieve the above objectives, the present invention provides the following solution:
[0006] Methods for estimating the surface dead combustible load in subtropical forests using multi-source remote sensing and machine learning include:
[0007] Acquire multi-source remote sensing data and auxiliary data within the target area and preprocess them; then acquire the preprocessed multi-source remote sensing data and auxiliary data.
[0008] Feature extraction is performed on preprocessed multi-source remote sensing data and auxiliary data to obtain feature variables;
[0009] Based on the aforementioned characteristic variables and the measured surface dead combustible load components, variables were selected, and a combustible load prediction model was constructed using multiple linear regression and machine learning models.
[0010] The combustible load prediction model was evaluated to obtain the optimal model. Among them, the XGBoost model was the optimal model for 1-hour time-delay combustible load, and the LightGBM model was the optimal model for litter load, 10-hour time-delay combustible load, 100-hour time-delay combustible load and total surface dead combustible load.
[0011] Using the optimal model, the surface dead combustible load of all vegetation types in the target area is regionally inverted to obtain the spatial distribution data of the surface dead combustible load.
[0012] Optionally, the multi-source remote sensing data includes Landsat 8 OLI optical images, Sentinel-1 C-band polarimetric synthetic aperture radar data, and ALOS-2 PALSAR-2 L-band polarimetric SAR data.
[0013] The auxiliary data includes land use data, canopy height data, and digital elevation model data.
[0014] Optionally, acquiring preprocessed multi-source remote sensing data and auxiliary data includes:
[0015] If missing or abnormal data is found in the multi-source remote sensing data and auxiliary data, it is filled in by interpolation method to determine the complete data set;
[0016] Based on the complete dataset, data format unification processing is performed to convert data from different sources into a consistent storage structure, resulting in a formatted dataset.
[0017] For the formatted dataset, data compression and storage optimization are performed, and a hierarchical storage method is adopted to determine the optimized dataset;
[0018] Vegetation index and texture features were calculated for Landsat 8 OLI optical images in the optimized dataset. Edge noise removal, thermal noise removal, and speckle filtering were performed on Sentinel-1 C-band polarimetric synthetic aperture radar data. Radiometric correction, terrain correction, and polarization decomposition were performed on ALOS-2 PALSAR-2 L-band polarimetric SAR data.
[0019] Optionally, edge noise removal, thermal noise removal, and speckle filtering processing are performed on Sentinel-1 C-band polarimetric synthetic aperture radar data, including:
[0020] The raw Sentinel-1 C image data is acquired, and edge noise is identified and cropped using a preset boundary detection method to obtain intermediate image data with reduced edge noise.
[0021] Based on the intermediate image data with reduced edge noise, a statistical analysis-based filtering method is used to suppress thermal noise, thereby obtaining intermediate image data after thermal noise processing.
[0022] By performing speckle filtering on the intermediate image data after thermal noise processing, and using the mean filtering method to smooth the image, smooth image data with reduced speckle noise is determined.
[0023] Based on the smoothed image data with reduced speckle noise, pixel-level contrast adjustment is performed on the signal quality of the radar image to obtain the adjusted image data.
[0024] For the adjusted image data, combined with the characteristics of polarization data, a dual-channel data polarization decomposition operation is performed to obtain polarized dual-polarization image data.
[0025] Optionally, feature extraction is performed on the preprocessed multi-source remote sensing data and auxiliary data to obtain feature variables, including: extracting feature variables from the preprocessed remote sensing data and auxiliary data based on the sampling point coordinates. The feature variables include vegetation index, texture features, polarization decibel (dB) value, tree height, vegetation type, slope, aspect, and altitude.
[0026] Optionally, variable selection based on the characteristic variables and the measured surface dead combustible load components includes:
[0027] For the multiple linear regression model, the stepwise regression method was adopted, with the Akaike Information Criterion as the screening criterion;
[0028] For machine learning models, the importance values of all variables are calculated separately, variables are added in descending order of importance, and the optimal combination of variables is determined based on the lowest calculated root mean square error.
[0029] Optionally, evaluating the combustible load prediction model to obtain the optimal model includes: evaluating the combustible load prediction model using leave-one-out cross-validation, and selecting the optimal model through the coefficient of determination, root mean square error, and mean absolute error.
[0030] Optionally, after obtaining the spatial distribution data of dead combustible load on the ground, the following may be included:
[0031] Based on the spatial distribution data of surface dead combustible load, the correlation between various types of vegetation and surface dead combustible load is corrected. If the load data of a certain type of vegetation deviates from the preset threshold, a second inversion is performed to obtain the corrected load data.
[0032] By combining the geometric characteristics of spatial distribution, the corrected load data is smoothed to obtain smoothed spatial distribution data of dead combustible load on the ground.
[0033] The beneficial effects of this invention are as follows: This invention combines multi-source remote sensing with machine learning to provide robust and spatially defined SDFL estimates, providing methodological support for fire risk assessment and precise combustible management in subtropical forests.
[0034] (1) Model comparisons show that machine learning models outperform linear regression models in SDFL inversion. Specifically, XGBoost was identified as the best model for 1-hour time-delay combustible load, while LightGBM achieved the highest accuracy in SDFL for litter load, 10-hour time-delay combustible load, 100-hour time-delay fuel combustible load, and total surface dead combustible load. Multiple linear regression and random forest models showed poor performance in all SDFL components.
[0035] (2) Variable importance analysis of the optimal model revealed a consistent pattern. Vegetation type was the most influential predictor of litter load, 10-hour time-delay combustible load, 100-hour time-delay combustible load, and total SDFL. For the 1-hour time-delay combustible load model, the Green-5-Me variable (the average of the green bands within a 5 × 5 window, derived from Landsat 8OLI texture features) showed the highest predictive power.
[0036] (3) Spatial inversion of SDFL using the optimal model for each combustible component reveals the obvious spatial heterogeneity of load distribution. Attached Figure Description
[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0038] Figure 1 This is a flowchart illustrating a method for estimating the surface dead combustible load of subtropical forests using multi-source remote sensing and machine learning, as described in an embodiment of the present invention. Detailed Implementation
[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0040] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0041] like Figure 1 As shown, this embodiment proposes a method for estimating the surface dead combustible load of subtropical forests using multi-source remote sensing and machine learning, including:
[0042] Acquire multi-source remote sensing data and auxiliary data within the target area and preprocess them; then acquire the preprocessed multi-source remote sensing data and auxiliary data.
[0043] Feature extraction is performed on preprocessed multi-source remote sensing data and auxiliary data to obtain feature variables;
[0044] Variables were selected based on characteristic variables and measured surface dead combustible load components, and a combustible load prediction model was constructed using multiple linear regression and machine learning models.
[0045] The combustible load prediction models were evaluated to obtain the optimal model. Among them, the XGBoost model was selected as the optimal model for 1-hour time-delay combustible load, and the LightGBM model was selected as the optimal model for litter load, 10-hour time-delay combustible load, 100-hour time-delay combustible load and total SDFL.
[0046] Using the optimal model, the surface dead combustible load of all vegetation types in the target area is regionally inverted to obtain the spatial distribution data of the surface dead combustible load.
[0047] Furthermore, the collection of measured surface dead fuel load components includes:
[0048] The focus of field sampling is to measure the surface non-combustibles of different vegetation types. The sampled vegetation types mainly include coniferous forests such as Pinus massoniana, Pinus elliottii, and Cunninghamia lanceolata (39 plots); broad-leaved forests dominated by Schima superba (23 plots); and coniferous and broad-leaved mixed forests (28 plots). For each plot, record the latitude, longitude, tree species composition, altitude, slope, diameter at breast height (DBH), tree height, and height to the first live branch.
[0049] To quantify the SDFL, three 2 m × 2 m sub-plots and three 1 m × 1 m sub-plots were established along the diagonal of each 20 m × 20 m plot, with intervals of 7 m, 14 m, and 21 m. In each 1 m × 1 m sub-plot, all fallen litter was collected and bagged. In the 2 m × 2 m sub-plots, dead fuels with 1-hour, 10-hour, and 100-hour time lags were collected and bagged separately.
[0050] After collecting the dead fuels, the samples were continuously oven-dried at 105 °C until a constant weight was reached, and the dry weight was measured to determine the dead fuel load. The interquartile range (IQR) method was used to remove outliers. This method defines the "reasonable range" as (Q1 - 1.5 × IQR, Q3 + 1.5IQR), where Q1 is the first quartile (25th percentile), Q3 is the third quartile (75th percentile), and IQR = Q3 - Q1. Values outside this range were determined to be outliers (low outliers: <Q1 - 1.5 × IQR; high outliers: >Q3 + 1.5 × IQR) and were excluded.
[0051] Furthermore, the multi-source remote sensing data includes Landsat 8 OLI optical images, Sentinel-1 C-band polarimetric synthetic aperture radar data, and ALOS-2 PALSAR-2 L-band polarimetric SAR data;
[0052] The auxiliary data includes land use data, canopy height data, and digital elevation model data.
[0053] Furthermore, obtaining the preprocessed multi-source remote sensing data and auxiliary data includes:
[0054] For the multi-source remote sensing data and auxiliary data, if data missing or anomalies are found, interpolation methods are used to fill them to determine a complete data set;
[0055] According to the complete data set, unified data format processing is implemented to convert data from different sources into a consistent storage structure to obtain a formatted data set;
[0056] For formatted datasets, implement data compression and storage optimization, adopt a hierarchical storage approach, and determine the optimal dataset;
[0057] Vegetation index and texture features were calculated for Landsat 8 OLI optical images in the optimized dataset. Edge noise removal, thermal noise removal, and speckle filtering were performed on Sentinel-1 C-band polarimetric synthetic aperture radar data. Radiometric correction, terrain correction, and polarization decomposition were performed on ALOS-2 PALSAR-2 L-band polarimetric SAR data.
[0058] Furthermore, edge noise removal, thermal noise removal, and speckle filtering are performed on Sentinel-1 C-band polarimetric synthetic aperture radar data, including:
[0059] The raw image data is obtained from Sentinel-1 C-band polarimetric synthetic aperture radar data. Edge noise is identified and cropped using a preset boundary detection method to obtain intermediate image data with reduced edge noise.
[0060] Based on the intermediate image data with reduced edge noise, a statistical analysis-based filtering method is used to suppress thermal noise, thereby obtaining intermediate image data after thermal noise processing.
[0061] By performing speckle filtering on the intermediate image data after thermal noise processing, and using the mean filtering method to smooth the image, smooth image data with reduced speckle noise is determined.
[0062] Based on the smoothed image data with reduced speckle noise, pixel-level contrast adjustment is performed on the signal quality of the radar image to obtain the adjusted image data.
[0063] For the adjusted image data, combined with the characteristics of polarization data, a dual-channel data polarization decomposition operation is performed to obtain polarized dual-polarization image data.
[0064] Radiometric correction, terrain correction, and polarization decomposition were performed on the ALOS-2PALSAR-2L band polarimetric synthetic aperture radar data to obtain dual-polarization imagery.
[0065] Furthermore, feature extraction is performed on the preprocessed multi-source remote sensing data and auxiliary data to obtain feature variables, including: based on the coordinates of sampling points, feature variables are extracted from the preprocessed remote sensing data and auxiliary data. Feature variables include vegetation index, texture features, polarization decibel (dB) value, tree height, vegetation type, slope, aspect, and altitude. Among them, the polarization decibel (dB) value is a logarithmic ratio index that combines polarization characteristics with decibel measurement method to characterize electromagnetic wave polarization-related parameters (such as polarization isolation, polarization loss, polarization purity, etc.).
[0066] Specifically, vegetation indices and texture features are key indicators for describing forest cover, as shown in Table 1. Vegetation indices primarily quantify vegetation growth and cover, while texture features capture the structure and spatial heterogeneity of the forest canopy. The Normalized Difference Vegetation Index (NDVI) is the most widely used index for monitoring vegetation cover, growth status, and phenological changes. The Enhanced Vegetation Index (EVI) more accurately reflects the growth changes of dense vegetation by reducing the signal saturation of atmospheric aerosols and high vegetation cover, improving the accuracy of vegetation dynamic monitoring. The FVC directly quantifies the spatial proportion of vegetation cover. The Green Vegetation Index (GVI) highlights the greenness of vegetation to distinguish between evergreen and deciduous species, while the Ratio Vegetation Index (RVI) is suitable for monitoring medium- to high-coverage vegetation and can quickly assess growth conditions. The Kernel Normalized Difference Vegetation Index (kNDVI) enhances the nonlinear difference between vegetation and the background, reduces soil and atmospheric interference, and is more sensitive to low-coverage vegetation. The Normalized Differential Moisture Index (NDMI) monitors vegetation moisture content, assesses drought stress, and indirectly reflects vegetation health. The Normalized Differential Water Index (NDWI) highlights water conditions and distinguishes between vegetation and water distribution. The Soil-Adjusted Vegetation Index (SAVI) accurately reflects sparse vegetation cover while minimizing the interference of soil reflectance on vegetation signals.
[0067] Table 1
[0068]
[0069] Gray-Level Co-occurrence Matrix (GLCM) texture features effectively describe the spatial structure of forests. Using the GLCM tool, texture metrics were calculated for four stride lengths (3×3, 5×5, 7×7, and 9×9), with different strides capturing different spatial scales of texture. A total of 192 spatial texture features were extracted, including mean (Me), variance (Var), homogeneity (Hom), contrast (Con), dissimilarity (Dis), entropy (Ent), angle-dichotomous matrix (ASM), and correlation (Cor). The mean reflects the overall gray-level in the vegetation image, indirectly corresponding to average vegetation cover, biomass, or growth vigor. Variance measures gray-level dispersion, indicating the uniformity of vegetation spatial distribution. Homogeneity quantifies the local consistency of vegetation texture, reflecting the structural similarity of vegetation between adjacent regions. Contrast measures the gray-level difference between adjacent vegetation regions, reflecting structural hierarchy or fragmentation. Dissimilarity describes the gray-level variation between adjacent pixels, highlighting surface roughness or structural complexity. Entropy measures the complexity and diversity of texture, reflecting the richness of vegetation type or structure. Angular second moment quantization of grayscale concentration indicates the uniformity and regularity of vegetation structure. Correlation assessment evaluates the linear relationship between adjacent vegetation pixels, reflecting the regularity or directionality of vegetation spatial distribution.
[0070] Both vegetation indices and texture features provide in-depth insights into canopy characteristics, which are closely correlated with SDFL. This correlation arises because surface dead combustible load is strongly influenced by canopy conditions, such as litter input rate and understory microclimate, which are mediated by canopy structure and activity. SAR polarization components are extracted as penetration structure features.
[0071] Vegetation type data were obtained from the 2021 Land Use Classification dataset, which classifies the study area into three forest types: broadleaf forest, coniferous forest, and mixed coniferous and broadleaf forest. Canopy height data were obtained from the Global Forest Canopy Height dataset. Topographic factors (elevation, slope, and aspect) also influence the distribution of SDFLs through synergistic interactions with vegetation. Variations in elevation, slope, and aspect affect the accumulation, spatial distribution, and transport of surface dead combustibles (SDFLs). Elevation, slope, and aspect were extracted from ASTER Global Digital Elevation Model (GDEM) data. Ancillary data factor variables are shown in Table 2.
[0072] Table 2
[0073]
[0074] Furthermore, variable selection based on characteristic variables and measured surface dead combustible load components includes:
[0075] For the multiple linear regression model, the stepwise regression method was adopted, with the Akaike Information Criterion as the screening criterion;
[0076] For machine learning models, the importance values of all variables are calculated separately, variables are added in descending order of importance, and the optimal combination of variables is determined based on the lowest calculated root mean square error.
[0077] Specifically, to evaluate the applicability of different models in deriving SDFL on complex terrain surfaces, multiple linear regression (a linear model) was compared with random forest, XGBoost, and LightGBM (a nonlinear ensemble model). The goal was to determine the most suitable model for SDFL estimation in the study area by evaluating the accuracy and stability of each model.
[0078] The multiple linear regression (MLR) model is used as a baseline linear model for comparison. It establishes a linear relationship between the dependent variable (dead combustible load) and multiple independent variables (remote sensing indices, texture features, and topographic variables) through an equation: + ;
[0079] In the formula As the dependent variable, The intercept is... As the independent variable, The coefficients are the values corresponding to the independent variables.
[0080] Random Forest (RF) Regression is an ensemble learning algorithm that constructs multiple decision trees and aggregates their predictions through majority voting or averaging. It randomly samples observations and features during tree construction, reducing overfitting and improving generalization performance. RF excels at handling high-dimensional data, naturally considering feature interactions and providing reliable ranking of feature importance. Its robustness to outliers and missing values makes it particularly suitable for heterogeneous data where dead combustible loads exhibit complex spatial patterns influenced by canopy structure, topography, and dead combustible decomposition.
[0081] Extreme Gradient Boosting (XGBoost) is an ensemble learning model based on the Gradient Boosting Decision Tree (GBDT) algorithm, belonging to a framework that combines additive models with iterative learning. This algorithm employs a greedy strategy to construct the decision tree: for each node, it evaluates all potential feature split points and selects the splitting method that maximizes the gain of the objective function.
[0082] XGBoost introduces several optimizations to the basic GBDT framework, delivering superior performance in terms of prediction accuracy, computational speed, and stability. Its core mechanism involves iteratively generating a series of decision trees and enhancing the overall predictive power of the model by weighting and combining their predictions.
[0083] Lightweight Gradient Boosting Machine (LightGBM) is a variant of Gradient Boosting Decision Tree (GBDT) algorithm, characterized by a breadth-first, horizontal tree growth strategy. In LightGBM, tree construction is performed by expanding nodes based on priority: from all current nodes, the node with the highest split gain is selected for splitting, and this process is repeated until a predefined tree depth threshold or the maximum number of leaf nodes is reached. This strategy focuses on optimizing high-gain nodes, thereby reducing redundant computation and improving training efficiency.
[0084] Given the large number of candidate variables (216 in total, including remote sensing features, vegetation indices, and topographic variables), multicollinearity among variables could reduce model accuracy. To address this issue, Stepwise Linear Regression (StepAIC) was employed to eliminate redundant variables and identify key explanatory variables. StepAIC uses both forward and backward selection methods to add variables that significantly contribute to the model while removing those that do not improve model accuracy. This process continues until no additional variables can enhance the model, and no further improvement is achieved by removing existing variables.
[0085] The Akaike Information Criterion (AIC) was used as the variable selection criterion; a lower AIC value indicates better model stability and fit performance. The variable combination corresponding to the minimum AIC value was selected as the final modeling variable set.
[0086] Feature selection is an effective strategy to reduce computational costs and improve model accuracy by focusing on the most influential variables. For the RandomForest model, the importance function in the R language's "RandomForest" package evaluates factor importance based on the IncMSE value; for the XGBoost model, the importance function in the "XGBoost" package uses the gain value to evaluate factor importance; and for the LightGBM model, the importance function in the "LightGBM" package also uses the gain value.
[0087] The initial set included a total of 216 variables. For each type of SDFL, the variable importance scores for each model were sorted in descending order. Model accuracy was then evaluated by progressively adding the most important variables. The optimal modeling variables were selected based on the importance ranking and the determined number of optimal variables.
[0088] Furthermore, the combustible load prediction model is evaluated to obtain the optimal model, including: using leave-one-out cross-validation to evaluate the combustible load prediction model, and selecting the optimal model through the coefficient of determination, root mean square error, and mean absolute error.
[0089] Specifically, both multiple linear regression and machine learning regression models are evaluated using Leave-One-Out Cross-Validation (LOOCV). In LOOCV, a sample is iteratively selected from the dataset as the test set, and the remaining samples form the training dataset. This process is repeated until every sample is used as the test set. The average of all individual test results is then used as a comprehensive evaluation metric for the model's performance.
[0090] LOOCV is particularly well-suited for small datasets because it maximizes the use of all data points, thereby minimizing bias. Furthermore, it offers high stability and reproducibility during evaluation. Model evaluation metrics include the coefficient of determination (R²). 2 R² is the root mean square error (RMSE) and mean absolute error (MAE). A higher R² indicates that the model fits the data better, while lower RMSE and MAE values indicate improved prediction accuracy.
[0091] Stepwise regression results showed that vegetation type was selected as the modeling variable for all SDFL components, including litter load, 1-hour time-delay combustible load, 10-hour time-delay combustible load, 100-hour time-delay combustible load, and total SDFL. Specifically, vegetation type was highly significantly correlated with litter load, 100-hour time-delay load, and total SDFL (P<0.001), and highly significantly correlated with 1-hour and 10-hour time-delay loads (P<0.05).
[0092] For litter combustible load, slope showed a highly significant correlation (p<0.001), while SWIR2-3-ASM and NIR-9-Me showed a very significant correlation (p<0.01). The 1-hour time-delay combustible load was significantly correlated with SWIR2-7-Me (p<0.05). The 10-hour time-delay fuel load was significantly correlated with Red-7-Cor (p<0.05). In contrast, the 100-hour time-delay load showed only weak correlations with all other variables, none reaching statistical significance (p>0.05).
[0093] Total SDFL was highly significantly correlated with slope and Red-3-ASM (p<0.01), and significantly correlated with SWIR1-9-Me and NIR-9-Me (p<0.05).
[0094] The optimal feature quantity selection based on machine learning determined the best-performing model for each SDFL component. LightGBM emerged as the best predictor for litter load, 10-hour (10-latency) combustible load, 100-hour (100-latency) combustible load, and total SDFL, while XGBoost was the best model for 1-hour (1-latency) combustible load. In contrast, Random Forest (RF) performed poorly across all combustible types.
[0095] The optimal number of features varies depending on the model and SDFL components: for litter load, 9 features are used for Random Forest, 24 features for XGBoost, and 24 features for LightGBM; for 1-hour time-delay fuel load, 4, 10, and 11 features are used respectively; for 10-hour time-delay fuel load, 26, 13, and 15 features are used; for 100-hour time-delay fuel load, 3, 14, and 13 features are used; and for total SDFL, 7, 13, and 18 features are used.
[0096] The variable selection results of the machine learning model show that vegetation type was consistently selected as a predictor for all SDFL components. It exhibited the highest importance in predicting litter load, 100-hour time-delay combustible load, and total SDFL; it reached peak importance for 1-hour time-delay combustible load in the LightGBM model; and it was also the variable with the greatest impact on 10-hour time-delay combustible load in both Random Forest and LightGBM.
[0097] For litter load, the other selected variables had relatively low importance across all three models. For 1-hour time-lag combustible load, Random Forest showed a balanced distribution of variable importance, while in XGBoost, variables such as Green-5-Me and NIR-5-Cor were more influential than others. In the LightGBM 1-hour time-lag combustible load model, vegetation type was far more important than all other variables. For 10-hour time-lag combustible load, Random Forest again showed a balanced distribution of variable importance, while vegetation type dominated in XGBoost and LightGBM, with other variables contributing the least.
[0098] Based on variable selection through stepwise regression, the MLR model for total SDFL achieved the highest accuracy, with an R² of 0.62 (RMSE = 0.94 t / ha, MAE = 0.75 t / ha). This was followed by the litter load model, with an R² of 0.52 (RMSE = 0.77 t / ha, MAE = 0.59 t / ha). However, the MLR models for time-delayed combustible loads performed poorly: the 1-hour time-delay model had an R² of 0.29 (RMSE = 0.14 t / ha, MAE = 0.11 t / ha); the 10-hour time-delay model had an R² of 0.27 (RMSE = 0.15 t / ha, MAE = 0.12 t / ha); and the 100-hour time-delay model had the lowest accuracy, with an R² of 0.17 (RMSE = 0.17 t / ha, MAE = 0.35 t / ha).
[0099] Machine learning models were built using variables selected through a feature selection process, and performance varied depending on the SDFL components and model type. In the random forest model, total SDFL and litter load showed relatively good performance: the R² for total SDFL was 0.53, RMSE was 1.04 t / ha, and MAEof was 0.83 t / ha, while the R² for litter load was 0.48, RMSE was 0.81 t / ha, and MAE was 0.62 t / ha. In contrast, the model performed poorly under time-delayed combustible loads, with R² values of 0.24, 0.12, and 0.25 for 1-hour, 10-hour, and 100-hour time-delayed combustible loads, respectively.
[0100] Similarly, the XGBoost model showed stronger performance in litter load (R²=0.51) and total SDFL (R²=0.60), but performed poorly in time-delayed combustibles, with R² values of 0.31 (1-hour delay), 0.06 (10-hour delay), and 0.24 (100-hour delay). Among the three machine learning models, the LightGBM model achieved the highest overall accuracy. It returned an R² value of 0.62 for litter load and 0.67 for total SDFL, while also outperforming Random Forest and XGBoost in both 10-hour and 100-hour time-delay combustible loads. Cross-model comparisons further showed that XGBoost performed best in 1-hour time-delay load, while LightGBM outperformed the other two models in all other SDFL components.
[0101] Notably, all three machine learning models consistently outperformed the multiple linear regression (MLR) model in terms of predictive accuracy. Therefore, for the subsequent regional inversion analysis, the XGBoost model was chosen for 1-time-delay combustible load estimation, and the LightGBM model was applied to all other SDFL components (leaf litter load, 10-time-delay combustible load, 100-time-delay combustible load, and total SDFL).
[0102] Different types of regional SDFLs are ranked according to their distribution proportions: the top 20% are Level 1, 20%-40% are Level 2, 40%-60% are Level 3, 60%-80% are Level 4, and 80%-100% are Level 5. The fire point distribution from January 2022 to July 2025 (using MODIS MOD14A1 product) is overlaid with the dead combustible load level. Then, the percentage of fire points occurring in each dead combustible load level is calculated.
[0103] Spatial mapping and assessment included ranking different types of regional SDFLs according to their distribution proportions: the top 20% were Level 1, 20%-40% were Level 2, 40%-60% were Level 3, 60%-80% were Level 4, and 80%-100% were Level 5. The distribution of fire points from January 2022 to July 2025 (using MODIS MOD14A1 products) was overlaid with the dead combustible load levels. Then, the percentage of fire points occurring in each dead combustible load level was calculated.
[0104] The results showed that: (1) Vegetation type is a key predictor, while canopy variables and topography contribute differently to each component; (2) Machine learning models outperformed MLR, with LightGBM achieving the highest accuracy in terms of litter load (R²=0.62), 10-time-lag combustible load (R²=0.29), 100-time-lag combustible load (R²=0.38), and total SDFL (R²=0.67), while XGBoost performed best in terms of 1-time-lag combustible load (R²=0.31); (3) Spatial mapping using the best models revealed significant heterogeneity, with high SDFL concentrated in coniferous forest areas. These results indicate that combining multi-source remote sensing with machine learning can provide robust and spatially defined SDFL estimates, offering methodological support for fire risk assessment and precise combustible management in subtropical forests.
[0105] Furthermore, after obtaining the spatial distribution data of dead combustible load on the ground, the following is included:
[0106] Based on the spatial distribution data of surface dead combustible load, the correlation between various types of vegetation and surface dead combustible load is corrected. If the load data of a certain type of vegetation deviates from the preset threshold, a second inversion is performed to obtain the corrected load data.
[0107] By combining the geometric characteristics of spatial distribution, the corrected load data is smoothed to obtain smoothed spatial distribution data of dead combustible load on the ground.
[0108] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for estimating the ground dead fuel load of subtropical forests using multi-source remote sensing and machine learning, characterized by, include: Acquire multi-source remote sensing data and auxiliary data within the target area and preprocess them; then acquire the preprocessed multi-source remote sensing data and auxiliary data. The multi-source remote sensing data includes Landsat 8 OLI optical images, Sentinel-1 C-band polarimetric synthetic aperture radar data, and ALOS-2 PALSAR-2 L-band polarimetric SAR data. The auxiliary data includes land use data, canopy height data, and digital elevation model data; Feature extraction is performed on preprocessed multi-source remote sensing data and auxiliary data to obtain feature variables, including: based on the coordinates of sampling points, feature variables are extracted from preprocessed remote sensing data and auxiliary data, and the feature variables include vegetation index, texture features, polarization decibel dB value, tree height, vegetation type, slope, aspect and altitude; Based on the aforementioned characteristic variables and the measured surface dead combustible load components, variables were selected, and a combustible load prediction model was constructed using multiple linear regression and machine learning models. Variable selection based on the aforementioned characteristic variables and measured surface dead combustible load components includes: For the multiple linear regression model, the stepwise regression method was adopted, with the Akaike Information Criterion as the screening criterion; For machine learning models, the importance values of all variables are calculated separately, variables are added in descending order of importance, and the optimal combination of variables is determined based on the lowest calculated root mean square error. The combustible load prediction model was evaluated to obtain the optimal model. Among them, the XGBoost model was the optimal model for 1-hour time-delay combustible load, and the LightGBM model was the optimal model for litter load, 10-hour time-delay combustible load, 100-hour time-delay combustible load and total surface dead combustible load. Using the optimal model, the surface dead combustible load of all vegetation types in the target area is regionally inverted to obtain the spatial distribution data of the surface dead combustible load.
2. The method of claim 1, wherein, Acquiring preprocessed multi-source remote sensing data and auxiliary data includes: If missing or abnormal data is found in the multi-source remote sensing data and auxiliary data, it is filled in by interpolation method to determine the complete data set; Based on the complete dataset, data format unification processing is performed to convert data from different sources into a consistent storage structure, resulting in a formatted dataset. For the formatted dataset, data compression and storage optimization are performed, and a hierarchical storage method is adopted to determine the optimized dataset; Vegetation index and texture features were calculated for Landsat 8 OLI optical images in the optimized dataset. Edge noise removal, thermal noise removal, and speckle filtering were performed on Sentinel-1 C-band polarimetric synthetic aperture radar data. Radiometric correction, terrain correction, and polarization decomposition were performed on ALOS-2PALSAR-2 L-band polarimetric SAR data.
3. The method of claim 2, wherein, Edge noise removal, thermal noise removal, and speckle filtering processing were performed on Sentinel-1 C-band polarimetric synthetic aperture radar data, including: The raw Sentinel-1 C image data is acquired, and edge noise is identified and cropped using a preset boundary detection method to obtain intermediate image data with reduced edge noise. Based on the intermediate image data with reduced edge noise, a statistical analysis-based filtering method is used to suppress thermal noise, thereby obtaining intermediate image data after thermal noise processing. By performing speckle filtering on the intermediate image data after thermal noise processing, and using the mean filtering method to smooth the image, smooth image data with reduced speckle noise is determined. Based on the smoothed image data with reduced speckle noise, pixel-level contrast adjustment is performed on the signal quality of the radar image to obtain the adjusted image data. For the adjusted image data, combined with the characteristics of polarization data, a dual-channel data polarization decomposition operation is performed to obtain polarized dual-polarization image data.
4. The method according to claim 1, characterized in that, The evaluation of the combustible load prediction model and the selection of the optimal model include: evaluating the combustible load prediction model using leave-one-out cross-validation, and selecting the optimal model through the coefficient of determination, root mean square error, and mean absolute error.
5. The method according to claim 1, characterized in that, After obtaining the spatial distribution data of dead combustible load on the ground, the following is included: Based on the spatial distribution data of surface dead combustible load, the correlation between various types of vegetation and surface dead combustible load is corrected. If the load data of a certain type of vegetation deviates from the preset threshold, a second inversion is performed to obtain the corrected load data. Combined with the geometric characteristics of spatial distribution, the corrected load data is smoothed to obtain smoothed spatial distribution data of surface dead combustible load.