A lfmc inversion method based on meteorological data time sequence characteristics
Patent Information
- Application Number
- CN202611010408.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-08
- Publication Date
- 2026-10-30
AI Technical Summary
[0005]尽管不同技术路径均已取得了一定的阶段性进展，但现有技术仍存在明显不足：一是气象驱动机制挖掘不足，多数研究仅使用同期气象数据，未系统揭示气象要素对LFMC的滞后效应与累积影响，难以反映植被水分的长期胁迫过程；二是时序特征利用不充分，缺乏多时间窗口、多统计量的系统性特征构建方法，高维特征冗余度高、筛选效率低；三是模型泛化能力有限，多数模型局限于特定区域或单一植被类型，缺少基于全球统一标准、大样本、多气候区的通用反演范式
本专利技术方案对传统基于气象数据的LFMC反演方法进行了改进，系统量化气象因子对LFMC的累积滞后效应，基于气象数据时间序列滞后特征实现了对LFMC的精准预测，为利用气象数据反演LFMC的相关研究提供了一种新思路。
Smart Images
Figure CN123012392A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing technology, and in particular to an LFMC (Live Fuel Moisture Content) inversion method based on the temporal characteristics of meteorological data. Background Technology
[0002] In recent years, global warming has intensified, and extreme weather events have become more frequent. Wildfires, as a sudden natural disaster facing the world, exhibit typical characteristics such as high frequency, wide range, and strong destructiveness. They have become a major disaster threatening the sustainable development of human society, and their harm has permeated multiple core areas such as life safety, economic development, and the ecological environment.
[0003] Live fuel moisture content (LFMC) is defined as the percentage of water mass to dry matter mass in living vegetation. It is a core physiological and ecological parameter characterizing the water balance within vegetation, assessing vegetation flammability, and predicting fire hazard levels and fire spread rates. LFMC directly determines the ease with which vegetation ignites, flame intensity, fire spread speed, and fire behavior patterns, and is a key indicator for global wildfire prevention and ecosystem health monitoring.
[0004] Traditional LFMC (Flame Risk Monitoring) acquisition relies on field sampling, drying, and weighing. While accurate, it suffers from drawbacks such as high cost, low efficiency, limited spatial coverage, and difficulty in long-term continuous observation, failing to meet the needs of large-area, near-real-time dynamic fire hazard monitoring. With the rapid development of remote sensing technology, its advantages of wide coverage, non-contact operation, and periodicity have gradually overcome the spatiotemporal limitations of traditional field measurements, becoming a core technology for dynamic monitoring of surface parameters. It can accurately acquire combustible material parameters over large areas, and LFMC inversion based on optical, thermal infrared, and microwave data is gradually becoming mainstream.
[0005] Although various technical approaches have made some progress, existing technologies still have significant shortcomings: First, the exploration of meteorological driving mechanisms is insufficient. Most studies only use contemporaneous meteorological data and fail to systematically reveal the lag effects and cumulative impacts of meteorological elements on LFMC, making it difficult to reflect the long-term stress process of vegetation water. Second, the utilization of temporal features is inadequate, lacking a systematic feature construction method with multiple time windows and multiple statistics, resulting in high redundancy and low screening efficiency of high-dimensional features. Third, the model's generalization ability is limited. Most models are limited to specific regions or single vegetation types, lacking a universal inversion paradigm based on globally unified standards, large samples, and multiple climate zones. Therefore, developing an LFMC inversion method that can quantify meteorological lag effects, efficiently screen features, has strong model generalization capabilities, and is applicable to multiple vegetation types globally has significant theoretical and applied value for improving the accuracy of wildfire early warning and perfecting eco-hydrological models. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention proposes an LFMC inversion method based on the time-series characteristics of meteorological data. By combining ERA5 meteorological data and measured data, it achieves accurate prediction of LFMC, providing a new approach for research on LFMC inversion using meteorological data. This method enables high-precision, wide-range, near-real-time, and predictable global LFMC inversion.
[0007] The specific technical solution of the present invention is as follows: An LFMC inversion method based on the time-series characteristics of meteorological data includes the following steps: Step 1: Acquire ERA5 meteorological data and LFMC measured data of active combustible material moisture content, and complete spatial matching of ERA5 meteorological data and LFMC measured data based on latitude and longitude information and sampling date; Step 2: Preprocess the meteorological data and LFMC measured data matched in Step 1, including: calculating derived data based on the original meteorological data and generating multidimensional meteorological factors; and performing vegetation type screening and sample filtering on the LFMC measured data to obtain a high-quality LFMC measured sample set. Step 3: For the sample set preprocessed in Step 2, extract multiple time-series statistical features for each meteorological factor in the meteorological data within different forward time windows of the LFMC sampling date, and construct high-dimensional time-series features; Step 4: For the high-dimensional time series features constructed in Step 3, the optimal lag window is selected by Pearson correlation coefficient, and then the optimal lag time series features for each meteorological factor in the meteorological data are obtained by screening through multicollinearity test and XGBoost feature importance average ranking. Step 5: Using the optimal meteorological data lag time series feature dataset selected in Step 4 as input, use another independent machine learning model XGBoost to perform LFMC inversion. The machine learning model XGBoost obtained after training is used as the LFMC estimation model, and LFMC inversion is performed based on the measured meteorological data.
[0008] Furthermore, the LFMC measured data product is Globe-LFMC 2.0, with a data time range from January 2000 to January 2023, and only data from countries with a sample size of more than 100 are selected; The meteorological data product is the ERA5-Land reanalysis dataset, which includes daily average data on temperature, dew point temperature, topsoil moisture content, and subsurface soil moisture content, as well as daily net solar radiation at 13:00, daily potential evaporation at 13:00, and daily precipitation at 0:00.
[0009] Furthermore, the meteorological factors include 2m temperature, relative humidity (RH), saturated vapor pressure difference (VPD), soil moisture content, net solar radiation at the surface, potential evaporation, and cumulative precipitation.
[0010] Furthermore, derived data is calculated from the original meteorological data to generate multiple types of meteorological factors, as detailed below: The relative humidity (RH) can be calculated from the raw data of 2m temperature and 2m dew point temperature. The mathematical expression is as follows:
[0011] in, Relative humidity, The temperature is 2m. The dew point temperature is 2m. The vapor pressure difference (VPD) is calculated from the raw data of 2m temperature and 2m dew point temperature. Its mathematical expression is:
[0012] in, This represents the saturated water vapor pressure difference.
[0013] Furthermore, the LFMC measured data undergoes vegetation type screening and sample filtering, as detailed below: Samples corresponding to the three vegetation types of trees, shrubs and herbs were retained, and samples with missing meteorological factor values were removed. At the same time, 3σ outlier removal was performed according to region and vegetation type. For a site that has multiple sets of samples of the same vegetation type at the same time, multiple sample values are taken as the final measured data of the site at that time.
[0014] Furthermore, step 3 is detailed as follows: For each LFMC sample sampling date, the forward time window was set to a length of 7, 15, 30, 60, 90, 120, 150, and 180 days respectively. For each time window, five time series statistical characteristics of seven meteorological factors were selected: mean, maximum, minimum, median, and linear trend.
[0015] Furthermore, step 4 is detailed as follows: For each meteorological factor, its corresponding time-series statistical characteristics under different forward time windows are calculated and Pearson correlation analysis is performed with the measured LFMC data. The correlation coefficient between each time-series statistical characteristic and the measured LFMC is calculated, and the time-series statistical characteristics are sorted from largest to smallest correlation, retaining the time window with the largest correlation. The statistical features obtained under different optimal time windows are filtered out to remove redundant features through multicollinearity test, and then the average value of feature importance is taken by the pre-trained XGBoost model to retain one optimal time series feature for each type of meteorological factor.
[0016] This invention provides an LFMC inversion method based on time-series features of meteorological data. First, ERA5 meteorological data and measured LFMC data are acquired and spatially matched to obtain a correlated dataset. The measured LFMC data is selected from global multi-location measured data provided by the Globe-LFMC 2.0 product; the ERA5 meteorological data uses the global 9km resolution diurnal ERA5-Land dataset product. Subsequently, the correlated dataset is preprocessed, and key derived data are calculated based on the original meteorological data. Considering the lagged and cumulative effects of meteorological factors on vegetation water content, the time-series features with the strongest correlation under the optimal time window for meteorological factors are extracted and screened. The screened correlated dataset is used as input, and the XGBoost model is used to achieve accurate LFMC prediction. Finally, the coefficient of determination (R²) and root mean square error (RMSE) are used to evaluate the inversion results.
[0017] Compared with the prior art, the present invention has the following beneficial effects: This patented technical solution improves the traditional LFMC inversion method based on meteorological data. It systematically quantifies the cumulative lag effect of meteorological factors on LFMC and achieves accurate prediction of LFMC based on the time series lag characteristics of meteorological data, providing a new approach for related research on LFMC inversion using meteorological data. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the LFMC inversion method based on the time-series characteristics of meteorological data.
[0019] Figure 2 Scatter plot of LFMC inversion accuracy for each vegetation type. Detailed Implementation
[0020] The present invention will now be described in detail with reference to the accompanying drawings and embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention. Example
[0021] like Figure 1 As shown in the figure, this embodiment provides an LFMC inversion method based on the time series characteristics of meteorological data, which includes the following steps: Step 1: Acquire ERA5 meteorological data and LFMC measured data. Based on the latitude and longitude information and sampling date of the LFMC measured samples, perform spatial matching between the ERA5 meteorological data and the LFMC measured data to obtain a correlated dataset. Specifically: The LFMC measured data used is the Globe-LFMC 2.0 product, which is currently the world's largest and most complete LFMC measured dataset, recording more than 280,000 LFMC measured data from 15 countries, spanning from 1977 to 2023, covering various ecosystems such as forests, shrublands, and grasslands. Based on this, considering the representativeness of each region and the sample size of the measured data, this example selects measured LFMC data from January 2000 to January 2023 from countries with large measured sample sizes, such as the United States, Spain, and France. The ERA5 meteorological data product is the ERA5-Land reanalysis dataset, which has an original spatial resolution of 9km and can provide daily hourly meteorological information on a global scale. It selects daily average data of temperature, dew point temperature, topsoil moisture content and subsurface soil moisture content, as well as net solar radiation at 13:00 every day, potential evaporation at 13:00 every day and total precipitation at 0:00 every day. The total precipitation at 0:00 every day can be used to characterize the cumulative precipitation of the previous day.
[0022] Step 2: Preprocess the meteorological data matched in Step 1 with the LFMC measured data. The preprocessing process includes: Step 2.1: Based on the raw meteorological data downloaded from the ERA5-Land reanalysis dataset, calculate the derived data. The implementation process includes: The relative humidity (RH) can be calculated from the raw data of 2m temperature and 2m dew point temperature. The mathematical expression is as follows:
[0023] in, Relative humidity, The temperature is 2m. The dew point temperature is 2m. The vapor pressure difference (VPD) is calculated from the raw data of 2m temperature and 2m dew point temperature. Its mathematical expression is:
[0024] in, This represents the saturated water vapor pressure difference.
[0025] The average of the surface soil moisture content and the subsurface soil moisture content is taken as the comprehensive soil moisture content. Combined with the derived calculation data, seven meteorological factors are finally retained: 2m temperature, relative humidity, VPD, soil moisture content, surface net solar radiation, potential evaporation, and cumulative precipitation.
[0026] Step 2.2: The LFMC measured data undergoes processing including vegetation type screening, outlier and missing value removal, duplicate sample merging, and sample size filtering to obtain a high-quality LFMC measured sample set. The process includes: Considering the limited sample size for some vegetation types, only samples corresponding to the three main vegetation types—trees, shrubs, and herbs—were retained. Samples with missing meteorological factor values were removed. Furthermore, 3σ outlier removal was performed based on region and vegetation type. Specifically: The samples were divided according to region and vegetation type, and the mean LFMC of the samples in each part was calculated. with standard deviation Remove Extreme values other than those mentioned above; For a site that has multiple sets of samples of the same vegetation type at the same time, multiple sample values are taken as the final measured data of the site at that time.
[0027] Step 3: For the sample set preprocessed in Step 2, extract various time-series statistical features for the seven meteorological factors within different forward time windows of the LFMC sampling day, and construct high-dimensional time-series features. Specifically: For each LFMC sample sampling date, the forward time window was set to a length of 7, 15, 30, 60, 90, 120, 150, and 180 days respectively. For each time window, five time series statistical characteristics of the seven meteorological factors were selected: mean, maximum, minimum, median, and linear trend.
[0028] Step 4: For the high-dimensional time-series features constructed in Step 3, the optimal lag window is selected using the Pearson correlation coefficient. Then, through multicollinearity testing and XGBoost feature importance averaging, the optimal meteorological data lag time-series features are selected. The specific implementation process is as follows: For each meteorological type, five time-series statistical characteristics are used to calculate the corresponding time-series statistical characteristics under different forward time windows and perform Pearson correlation analysis with the measured LFMC data. The correlation coefficient between each time-series statistical characteristic and the measured LFMC is calculated, and the time-series statistical characteristics are sorted from largest to smallest correlation, retaining the time window with the largest correlation. The statistical features obtained under different optimal time windows are filtered out to remove redundant features through multicollinearity test. Then, the pre-trained XGBoost model is trained multiple times to obtain the average value of feature importance, and one optimal time series feature is retained for each type of meteorology.
[0029] Step 5: Using the optimal meteorological data lag time series feature dataset selected in Step 4 as input, use another independent machine learning model XGBoost to perform LFMC inversion. The machine learning model XGBoost obtained after training is used as the LFMC estimation model, and LFMC inversion is performed based on the measured meteorological data.
[0030] The XGBoost model is renowned for its superior computational efficiency and flexible applicability, and is widely used in tasks such as ranking, classification, and regression. XGBoost improves and optimizes the Gradient Boosting Framework. Its core idea is to combine multiple weak learners (usually decision trees) into a strong learner, thereby improving the model's predictive performance and meeting the requirements.
[0031] Step 6: Evaluate the accuracy of the LFMC inversion obtained in Step 5: This example uses the coefficient of determination. ) and root mean square error (RMSE) Two evaluation indicators are used to evaluate the inversion effect of LFMC, and the calculation formulas are as follows:
[0032]
[0033] in, The actual measured value. These are the model's predicted values. is the mean of the true values, and n is the number of samples.
[0034] Evaluation results as follows Figure 2 As shown, it can be seen that by extracting and filtering the time series characteristics of meteorological data, a high-precision LFMC estimation can be achieved, and this inversion method has certain applicability under different vegetation types.
[0035] This embodiment provides an LFMC inversion method based on the time series characteristics of meteorological data. The data used includes measured LFMC data from multiple locations worldwide from Globe-LFMC 2.0 and hourly meteorological data at a global resolution of 9 km based on the ERA5-Land reanalysis dataset. Through systematic processing and feature engineering of the meteorological data, a dataset is constructed, and an XGBoost model is used to achieve accurate LFMC prediction. Finally, the coefficient of determination R0 is used... 2 The root mean square error (RMSE) was used to evaluate the inversion accuracy. The feasibility of using meteorological data to invert LFMC was explored, and the results showed acceptable accuracy.
[0036] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.
Claims
1. An LFMC inversion method based on the time-series characteristics of meteorological data, characterized in that, Includes the following steps: Step 1: Acquire ERA5 meteorological data and LFMC measured data of active combustible material moisture content, and complete spatial matching of ERA5 meteorological data and LFMC measured data based on latitude and longitude information and sampling date; Step 2: Preprocess the meteorological data and LFMC measured data matched in Step 1, including: calculating derived data based on the original meteorological data and generating multidimensional meteorological factors; and performing vegetation type screening and sample filtering on the LFMC measured data to obtain a high-quality LFMC measured sample set. Step 3: For the sample set preprocessed in Step 2, extract multiple time-series statistical features for each meteorological factor in the meteorological data within different forward time windows of the LFMC sampling date, and construct high-dimensional time-series features; Step 4: For the high-dimensional time series features constructed in Step 3, the optimal lag window is selected by Pearson correlation coefficient, and then the optimal lag time series features for each meteorological factor in the meteorological data are obtained by screening through multicollinearity test and XGBoost feature importance average ranking. Step 5: Using the optimal meteorological data lag time series feature dataset selected in Step 4 as input, use another independent machine learning model XGBoost to perform LFMC inversion. The machine learning model XGBoost obtained after training is used as the LFMC estimation model, and LFMC inversion is performed based on the measured meteorological data.
2. The LFMC inversion method based on the time-series characteristics of meteorological data according to claim 1, characterized in that, The LFMC measured data product is Globe-LFMC 2.0, with a data time range from January 2000 to January 2023, and only data from countries with more than 100 samples are selected; The meteorological data product is the ERA5-Land reanalysis dataset, which includes daily average data on temperature, dew point temperature, topsoil moisture content, and subsurface soil moisture content, as well as daily net solar radiation at 13:00, daily potential evaporation at 13:00, and daily precipitation at 0:
00.
3. The LFMC inversion method based on the time-series characteristics of meteorological data according to claim 1, characterized in that, The meteorological factors include 2m temperature, relative humidity (RH), saturated vapor pressure difference (VPD), soil moisture content, net solar radiation at the surface, potential evaporation, and cumulative precipitation.
4. The LFMC inversion method based on the time-series characteristics of meteorological data as described in claim 3, characterized in that, The process involves calculating derived data from the original meteorological data to generate multiple types of meteorological factors, as detailed below: The relative humidity (RH) can be calculated from the raw data of 2m temperature and 2m dew point temperature. The mathematical expression is as follows: ， in, Relative humidity, The temperature is 2m. The dew point temperature is 2m. The vapor pressure difference (VPD) is calculated from the raw data of 2m temperature and 2m dew point temperature. Its mathematical expression is: ， in, This represents the saturated water vapor pressure difference.
5. The LFMC inversion method based on the time-series characteristics of meteorological data according to claim 1, characterized in that, The vegetation type screening and sample filtering processes performed on the LFMC measured data are as follows: Samples corresponding to the three vegetation types of trees, shrubs and herbs were retained, and samples with missing meteorological factor values were removed. At the same time, 3σ outlier removal was performed according to region and vegetation type. For a site that has multiple sets of samples of the same vegetation type at the same time, multiple sample values are taken as the final measured data of the site at that time.
6. The LFMC inversion method based on the time-series characteristics of meteorological data according to claim 1, characterized in that, Step 3 is described in detail below: For each LFMC sample sampling date, the forward time window was set to a length of 7, 15, 30, 60, 90, 120, 150, and 180 days respectively. For each time window, five time series statistical characteristics of seven meteorological factors were selected: mean, maximum, minimum, median, and linear trend.
7. The LFMC inversion method based on the time-series characteristics of meteorological data according to claim 6, characterized in that, Step 4 is described in detail below: For each meteorological factor, its corresponding time-series statistical characteristics under different forward time windows are calculated and Pearson correlation analysis is performed with the measured LFMC data. The correlation coefficient between each time-series statistical characteristic and the measured LFMC is calculated, and the time-series statistical characteristics are sorted from largest to smallest correlation, retaining the time window with the largest correlation. The statistical features obtained under different optimal time windows are filtered out to remove redundant features through multicollinearity test, and then the average value of feature importance is taken by the pre-trained XGBoost model to retain one optimal time series feature for each type of meteorological factor.