Method and system for estimating vegetation canopy fuel moisture content based on meteorological and remote sensing data

Abstract

The application discloses a kind of estimation method and system of vegetation canopy combustible moisture content based on meteorology and remote sensing data, comprising the following steps: first, combustible moisture content and various meteorological data and remote sensing data and other several kinds of subsidiary data are selected as combustible moisture content estimation data.Meteorological data includes air temperature, relative humidity, rainfall and wind speed.Remote sensing data includes two vegetation indexes: enhanced vegetation index and normalized vegetation index.Subsidiary data includes: root zone soil moisture, vapor pressure difference, drought index, fire weather factor.Then the long time sequence characteristics of meteorological data are extracted.The size of time window is determined first, and the experimental results show that the correlation coefficient of most sites is relatively high under the time window of 90-210 days.90 days, 150 days and 210 days of time window are selected respectively to extract the time characteristics of four kinds of meteorological data.Secondly, the samples of experimental area are divided into five vegetation classifications, which are closed shrub, sparse shrub, multi-tree tropical grassland, tropical savanna and grassland.Finally, the data sets of the five different vegetation types are sequentially adjusted to obtain the respective estimation model.

Description

Technical Field

[0001] This invention relates to the field of remote sensing technology, and in particular to a method for estimating the moisture content of combustibles in vegetation canopy based on meteorological and remote sensing data. Background Technology

[0002] Wildfires cause severe damage to ecosystems: once a wildfire breaks out, it not only destroys almost all surrounding flora and fauna, but also endangers the safety of nearby residents and rescue workers due to its difficulty in control. Therefore, the only way to prevent wildfires is to take timely measures before they occur. Thus, providing early and accurate warnings for wildfires is a key focus of current research. The moisture content of vegetation canopy fuels is closely related to the occurrence and spread of wildfires; the moisture content of fuels is a major determinant of the ignition and spread speed of fires.

[0003] Currently, there are several methods for obtaining Live Fuel Moisture Content (LFMC) data. Field sampling is the most direct and simplest method, but due to its complexity and the significant manpower, material resources, and time required, it is difficult to implement on a regional or global scale. Many fire risk assessment and early warning systems currently use meteorological indices to estimate fuel moisture content, but these are primarily used for estimating the moisture content of dead fuels. This is because their relationship with atmospheric conditions is more direct than that of live fuels, and they are easier to ignite. Dead fuels, especially fine fuels, respond rapidly to atmospheric conditions and can be modeled based on inputs such as temperature, humidity, or insufficient vapor pressure. However, in contrast, Live Fuel Moisture Content (LFMC) is difficult to model because moisture content is closely related to plant physiological state, and there can be significant differences between different species and seasonal variations. Living plants control their stomatal conductance under drought conditions and have the ability to extract water from the soil; therefore, the relationship between meteorological indices and the moisture content of living vegetation is less direct than that of dead vegetation. Furthermore, due to the low spatial resolution of meteorological data, most require further interpolation to obtain a comprehensive spatial-scale estimate of the water content of combustibles. Although applying meteorological indicators to living plants presents challenges, several drought indicators have been correlated with LFMC in the Mediterranean environment, showing good consistency.

[0004] Therefore, applying meteorological data to the estimation of moisture content in combustible vegetation has research value and development potential. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data. This method mainly utilizes meteorological and remote sensing data to mine time features related to LFMC from meteorological data and uses the XGBoost machine learning method to estimate the moisture content of vegetation canopy.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The method for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data provided by this invention includes the following steps: The moisture content data of combustibles, meteorological data, remote sensing data, and auxiliary data are acquired. The moisture content data of combustibles is field sampling data, and the meteorological data and remote sensing data are used as basic feature variables. Long-term time series feature extraction of meteorological data; Construct a database to store datasets of different vegetation cover types; An XGBoost model was constructed, and the dataset was divided into training and validation sets according to vegetation cover type. The parameters of the XGBoost model were tuned to obtain the optimal model, which yielded the optimal model for estimating the moisture content of combustibles. The validation set is input into the estimation model for optimal combustible moisture content to obtain the FMC prediction results.

[0007] Furthermore, the remote sensing data comprises four vegetation indices: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EDI), Visible Atmospheric Drag Index (VAD), and Normalized Infrared Index (NIIR). The meteorological data includes wind speed, temperature, rainfall, and relative humidity. The auxiliary data includes root zone soil moisture, vapor pressure difference, drought index, and fire weather factors.

[0008] Furthermore, it also includes the following steps: Determine the size of the time window for extracting the temporal features of long-term meteorological data. For each piece of combustible material moisture content (LFMC) data, calculate the moving average of the corresponding meteorological data within a preset time range, and calculate the correlation coefficient. Determine the interval of the correlation coefficient within a preset value as the time window selection interval.

[0009] Furthermore, the long-term meteorological data time series features include 90-day, 150-day, and 210-day time feature factors of the meteorological data; the vegetation cover type database is constructed based on the IGBP classification of MCD12Q1.

[0010] Furthermore, the XGBoost model is constructed in the following manner: XGBoost is an additive model consisting of k base models. Assume the tree model to be trained in the t-th iteration is... Then we have:

[0011] in, This represents the prediction result for sample i after the t-th iteration. The prediction results are for the first t-1 trees. Let be the model for the t-th tree; A model representing the k-th tree; Furthermore, the loss function in the XGBoost model is calculated according to the following formula:

[0012] in, This represents the loss function for all samples; n is the number of samples. The loss function representing a single sample; Indicates predicted value Represents the actual value.

[0013] Furthermore, the objective function in the XGBoost model is defined as follows:

[0014] in, Represents leaf nodes The sum of the first-order partial derivatives of all included samples; Represents leaf nodes The sum of the second-order partial derivatives of the included samples; This represents the j-th leaf node; Indicates the number of leaf nodes; For regularization parameters; This represents the weight of the j-th leaf node.

[0015] Furthermore, the XGBoost model is trained in the following manner: When building the t-th tree, a greedy algorithm is used to calculate the optimal split point of the leaf node; Calculate the difference in structural scores before and after branching, denoted as Gain. Select the branching point on the feature point with the largest Gain for branching. When Gain is less than the minimum value set in the model, the tree stops growing. Assuming feature splitting is completed at a certain node, the objective function before splitting can be written as:

[0016] The objective function after splitting is:

[0017] Therefore, for the objective function, the benefit after splitting is:

[0018] This represents the left leaf node that splits from the original leaf node. The sum of the first-order partial derivatives of all included samples; This represents the right leaf node that splits from the original leaf node. The sum of the first-order partial derivatives of all included samples; This represents the left leaf node that splits from the original leaf node. The sum of the second-order partial derivatives of the included samples; This represents the right leaf node that splits from the original leaf node. The sum of the second-order partial derivatives of the included samples.

[0019] The present invention provides a system for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above method.

[0020] The beneficial effects of this invention are as follows: This invention provides a method for estimating the LFMC (Fuel Content of Vegetation Canopy) based on meteorological and remote sensing data. The model integrates meteorological and remote sensing data, and its greatest advantage lies in its primary use of meteorological data and the extraction of long-term time-series features related to LFMC from this data, thereby improving the estimation accuracy. Vegetation moisture content affects the physiological condition of vegetation, and meteorological conditions are closely related to the physiological state of vegetation. However, this correlation is not immediate but accumulates through long-term effects. Therefore, long-term time-series feature factors are the focus and key of this method. Through analysis and mining, the extracted long-term time-series features exhibit significant stability, all being long-term moving averages. These features are simple to calculate and have clear physical meaning, significantly improving the model's accuracy while avoiding increased complexity.

[0021] Meanwhile, this method achieved good accuracy even when the selected experimental area was large, the climate types varied, and the sample distribution was sparse. It also divided the data into multiple vegetation types, analyzed the estimation effect under each vegetation type, and achieved good results, indicating that the model has high universality.

[0022] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0023] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 The flowchart shows the LFMC estimation method.

[0024] Figure 2 This is a schematic diagram of the gradient boosting tree.

[0025] Figure 3 This is a flowchart of a method for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and specific 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.

[0027] Example 1

[0028] like Figure 1 As shown in this embodiment, the method for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data includes the following steps: The moisture content data of combustibles, meteorological data, remote sensing data, and auxiliary data are acquired. The moisture content data of combustibles is field sampling data, and the meteorological data and remote sensing data are used as basic feature variables. Long-term time series features were extracted from meteorological and soil moisture data. The extracted feature factors, remote sensing data, and auxiliary data were used as input factors of the model, and combustible material moisture content data were used as dependent variable data of the model. A database is constructed, which includes vegetation cover types, namely, five types of data: dense shrubland, sparse shrubland, multi-tree tropical grassland, tropical savanna and grassland. An XGBoost model is constructed, and a training set and a validation set are determined. The training set and the validation set include five vegetation cover types. The parameters of the XGBoost model are tuned to obtain the optimal model, and the optimal model for estimating the moisture content of combustibles is obtained.

[0029] In this embodiment, the remote sensing data consists of four vegetation indices: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EDI), Visible Atmospheric Drag Index (VAD), and Normalized Infrared Index (NIR). The meteorological data includes wind speed, temperature, rainfall, and relative humidity. The auxiliary data includes root zone soil moisture, vapor pressure difference, drought index, and fire weather factors.

[0030] In this embodiment, the size of the time window for extracting the time features of long-term meteorological data is determined. For each piece of combustible material moisture content LFMC data, the moving average of the corresponding meteorological data within a preset time range (1-360 days) is calculated, and its correlation coefficient is calculated. The interval with the correlation coefficient at a preset value (higher) is determined as the time window selection interval. The long-term meteorological data time series features described in this embodiment include 90-day, 150-day, and 210-day time feature factors of the meteorological data; The vegetation cover type database described in this embodiment is constructed based on the IGBP classification of MCD12Q1; The XGBoost model described in this embodiment uses the gradient boosting algorithm to iterate with trees. Each time a new tree is added, the residual of the previous prediction is fitted, and finally k decision trees are obtained. The sample to be predicted is placed into the leaf node corresponding to each tree, and the scores of the leaf nodes are added together to get the predicted value.

[0031] This embodiment provides a method for estimating the moisture content of vegetation combustibles using the XGBoost machine learning model. XGBoost, short for eXtreme Gradient Boosting, is an efficient system implementation of Gradient Boosting. XGBoost is a tool for massively parallel boosting trees; it is an optimized distributed gradient boosting library that is efficient, flexible, and portable. Figure 2 This demonstrates the principle of gradient boosting trees.

[0032] XGBoost is an additive model consisting of k base models. Assume the tree model to be trained in the t-th iteration is... Then we have:

[0033] in, This represents the prediction result for sample i after the t-th iteration. The prediction results are for the first t-1 trees. Let be the model for the t-th tree; A model representing the k-th tree; The loss function in the XGBoost model can be derived from the predicted values. Compared with the true value Indicate:

[0034] in, This represents the loss function for all samples, where n is the number of samples. The loss function representing a single sample; The prediction accuracy of a model is determined by both its bias and variance. The loss function represents the model's bias. To minimize variance, a regularization term is added to the objective function to prevent overfitting. Therefore, the objective function is derived from the model's loss function. Together with the regularization term Ω to suppress model complexity, the objective function is defined as follows:

[0035] in, It sums the complexity of all t trees and adds it to the objective function as a regularization term to prevent the model from overfitting. Represent the objective function; Since XGBoost is an algorithm in the boosting family, it follows a forward step-by-step addition method. Taking the model in the i-th step as an example, the model iterates over the i-th sample... The predicted value is:

[0036] in, It is the first The predicted values ​​given by the step model are known constants. These are the predicted values ​​from the new model that needs to be added this time. At this point, the objective function can be written as:

[0037] Taylor expansion:

[0038] Define a decision tree, which consists of two parts: leaf node weight vector ; Mapping relationship from instance (sample) to leaf node q (essentially a branching structure of a tree);

[0039] The complexity of decision trees It can be composed of the number of leaf nodes T. The fewer the leaf nodes, the simpler the model. In addition, leaf nodes should not contain excessively high weights. (Analogous to the weights of each variable in LR), so the regularization term of the objective function is a vector consisting of the number of leaf nodes in all generated decision trees and the weights of all nodes. Paradigm is jointly determined.

[0040]

[0041] All samples x belonging to the j-th leaf node i The sample set that is added to a leaf node can be mathematically represented as: Therefore, the objective function of the XGBoost model can be written as:

[0042] To simplify the expression, define The meaning is as follows: Represents leaf nodes The sum of the first-order partial derivatives of all included samples is a constant. Represents leaf nodes The sum of the second-order partial derivatives of the included samples is a constant; Will and Substituting this into the objective function of the XGBoost model, the final objective function is:

[0043] When the subexpression of each leaf node reaches the maximum value point, that is... The entire objective function then reaches its maximum / minimum point.

[0044] The objective function can now be simplified to:

[0045] in, This represents the j-th leaf node; Indicates the number of leaf nodes; For regularization parameters; During training, when building the t-th tree, a crucial problem is finding the optimal split point for the leaf nodes. A greedy algorithm is used for this purpose. The essence of the greedy algorithm is to use an objective function (structure function) to define the quality of the structure, calculate the difference in structure score before and after branching, denoted as Gain, select the branch point on the feature point with the largest Gain for branching, and stop the tree from growing when Gain is less than the minimum value set by the model.

[0046] Assuming feature splitting is completed at a certain node, the objective function before splitting can be written as:

[0047] The objective function after splitting is:

[0048] Therefore, for the objective function, the benefit after splitting is:

[0049] This represents the left leaf node that splits from the original leaf node. The sum of the first-order partial derivatives of all included samples; This represents the right leaf node that splits from the original leaf node. The sum of the first-order partial derivatives of all included samples; This represents the left leaf node that splits from the original leaf node. The sum of the second-order partial derivatives of the included samples; This represents the right leaf node that splits from the original leaf node. The sum of the second-order partial derivatives of the included samples; For regularization parameters; The greedy algorithm's calculation process starts from a tree depth of 0: Enumerate all available features for each leaf node; For each feature, the training samples belonging to that node are sorted in ascending order according to the feature value. The best split point for that feature is determined by linear scanning, and the splitting gain of that feature is recorded. Select the feature with the highest benefit as the splitting feature, use the best splitting point of the feature as the splitting position, split the node into two new leaf nodes, and associate each new node with the corresponding sample set. Return to step 1 and recursively execute until a specific condition is met; The following section details XGBoost with data. XGBoost achieves the fitting process by continuously building a tree. To reduce model complexity, input factors were selected, with a total of 11 factors added to the model. Simultaneously, to obtain the optimal XGBoost model for estimating LFMC and to further reduce model complexity, a detailed parameter tuning process was performed, adjusting different parameter configurations to achieve the minimum model RMSE. Tuner parameters included iteration step size, minimum loss function decrease required for node splitting, maximum tree depth, column sampling ratio for each node split, random sampling ratio, and minimum sum of leaf node sample weights. Convergence occurred when the XGBoost tree reached 200. The model RMSE was minimized when the iteration step size was 0.1, the minimum loss function decrease required for node splitting was 0.3, the maximum tree depth was 10, the column sampling ratio for each node split was 80%, the random sampling ratio was 85%, and the minimum sum of leaf node sample weights was 4.

[0050] When building a new tree, firstly, 85% of the samples with replacement are drawn from all samples. Then, only 80% of the features are extracted from the drawn samples for building the current tree. For the root node, which is where all sampled samples fall, all available features for the current node are enumerated. For each feature, such as ndvi, all samples are first sorted according to their ndvi values ​​from smallest to largest. Since its distribution range is [-1, 1], split points are sequentially defined within this range. Then, the splitting gain at different split points is calculated and recorded. Finally, this process is performed sequentially for all 11 features. Finally, the split point with the best splitting gain is selected from all split points across all features as the split point for that node. All nodes are split sequentially until the splitting gain is less than the complexity penalty term. If the tree depth exceeds 10, the splitting stops, the tree is complete, and new trees are rebuilt until 200 trees are built, at which point the model is complete.

[0051] Example 2

[0052] like Figure 3 As shown in this embodiment, the method for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data specifically includes the following steps: (1) Preparation of raw data The target for estimation is Live Fuel Moisture Content (LFMC), which is the ratio of the difference between the wet weight and the dry weight of the plant in the sampling unit to the dry weight. It reflects the water content of a unit of vegetation leaf. The data used in this example are obtained from field sampling.

[0053] In this embodiment, combustible material moisture content, along with various meteorological and remote sensing data, and several other supplementary data, are selected as the data for estimating combustible material moisture content. Meteorological data includes air temperature, relative humidity, rainfall, and wind speed. Remote sensing data includes two vegetation indices: enhanced vegetation index and normalized difference vegetation index (NDVI). Supplementary data includes: root zone soil moisture, vapor pressure difference, drought index, and fire weather factors.

[0054] (2) Extraction of time series features Long-term series features were extracted from meteorological data. First, the time window size was determined. For each LFMC data point, the 1-360-day moving average of its corresponding meteorological data was calculated, and its correlation coefficient was determined to identify the intervals with higher correlation coefficients. Experimental results showed that the correlation coefficients of most stations increased within a 90-210 day time window. Therefore, 90-day, 150-day, and 210-day time windows were selected to extract the temporal features of the four types of meteorological data. The results showed significant consistency: under different time windows, the moving average sum of the four types of meteorological data was the most correlated factor with LFMC.

[0055] (3) Classification of vegetation types Based on the IGBP classification model of MCD12Q1, the samples in the experimental area were divided into 5 vegetation categories: dense shrubland, sparse shrubland, sparse tropical grassland, tropical savanna, and grassland.

[0056] (4) Training the model and analyzing the FMC prediction results After comparing a series of machine learning models, XGBoost was ultimately selected. Data sets from five different vegetation types were sequentially input into the model to obtain their respective accuracies. Then, the accuracy of all samples in the model was calculated. The model results are shown in Table 1, which presents the water content estimation results for each vegetation type. Specifically, the R-value for dense shrubland is... 2 R reaches 0.63 for sparse shrubs 2 Reaching 0.56, the R of the tropical savanna with many trees 2 Reaching 0.42, the R of tropical savanna 2 Reaching 0.43 and R of grassland 2 The R-value reached 0.55. The R-value for all samples... 2 It reached 0.53.

[0057] Table 1

[0058] This embodiment provides a method for estimating the LFMC (Fuel Content of Vegetation Canopy Combustible Material) based on meteorological and remote sensing data. By extracting long-term time-series features from meteorological data and incorporating them into the model, along with remote sensing data and other supplementary data, the accuracy of LFMC estimation is improved, demonstrating that the long-term time characteristics of meteorological data are beneficial for LFMC estimation.

[0059] 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. A method for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data, characterized in that: Includes the following steps: The moisture content data of combustibles, meteorological data, remote sensing data, and auxiliary data are acquired. The moisture content data of combustibles is field sampling data, and the meteorological data and remote sensing data are used as basic feature variables. Long-term time series feature extraction of meteorological data; Determine the size of the time window for long-term time series feature extraction. For each LFMC data point on combustible material moisture content, calculate the moving average of the corresponding meteorological data within a preset time range, and calculate the correlation coefficient. Determine the interval of the correlation coefficient within a preset value as the time window selection interval. The long-term time series features include 90-day, 150-day, and 210-day time feature factors of meteorological data; Construct a database to store datasets of different vegetation cover types; The database of vegetation cover types was constructed based on the IGBP classification of MCD12Q1; An XGBoost model was constructed, and the dataset was divided into training and validation sets according to vegetation cover type. The parameters of the XGBoost model were tuned to obtain the optimal model, which yielded the optimal model for estimating the moisture content of combustibles. The validation set is input into the estimation model for optimal combustible moisture content to obtain the FMC prediction results.

2. The method for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data as described in claim 1, characterized in that: The remote sensing data consists of four vegetation indices: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EDI), Visible Atmospheric Drag Index (VAD), and Normalized Infrared Index (NIR). The meteorological data includes wind speed, temperature, rainfall, and relative humidity. The auxiliary data includes root zone soil moisture, vapor pressure difference, drought index, and fire weather factor.

3. The method for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data as described in claim 1, characterized in that: The XGBoost model is constructed in the following manner: XGBoost is an additive model consisting of k base models. Assume the tree model to be trained in the t-th iteration is... Then we have: in, This represents the prediction result for sample i after the t-th iteration. This is the prediction result for the (t-1)th iteration. Let be the model for the t-th iteration; Let represent the model in the k-th iteration.

4. The method for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data as described in claim 1, characterized in that: The loss function in the XGBoost model is calculated according to the following formula: in, This represents the loss function for all samples; n is the number of samples. The loss function representing a single sample; Indicates predicted value Represents the actual value.

5. The method for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data as described in claim 1, characterized in that: The objective function in the XGBoost model is defined as follows: in, Represents leaf nodes The sum of the first-order partial derivatives of all included samples; Represents leaf nodes The sum of the second-order partial derivatives of the included samples; This represents the j-th leaf node; Indicates the number of leaf nodes; For regularization parameters; This represents the weight of the j-th leaf node.

6. The method for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data as described in claim 5, characterized in that: The XGBoost model is trained in the following manner: When establishing the t-th iteration, a greedy algorithm is used to calculate the optimal split point of the leaf node; Calculate the difference in structural scores before and after branching, denoted as Gain. Select the branching point on the feature point with the largest Gain for branching. When Gain is less than the minimum value set in the model, the tree stops growing. Assuming feature splitting is completed at a certain node, the objective function before splitting can be written as: The objective function after splitting is: Therefore, for the objective function, the benefit after splitting is: This represents the left leaf node that splits from the leaf node. The sum of the first-order partial derivatives of all included samples; This represents the right leaf node that splits from the leaf node. The sum of the first-order partial derivatives of all included samples; This represents the left leaf node that splits from the leaf node. The sum of the second-order partial derivatives of the included samples; This represents the right leaf node that splits from the leaf node. The sum of the second-order partial derivatives of the included samples.

7. A system for estimating the moisture content of vegetation canopy combustibles based on meteorological and remote sensing data, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the method described in any one of claims 1 to 6.

Images