A method for predicting vegetation fuel moisture content based on a space-time interpolation method

Abstract

The application relates to a vegetation combustible moisture content prediction method based on a space-time interpolation method, which comprises the following steps: S1, data acquisition and preprocessing S11, vegetation time sequence FMC data acquisition S12, invalid data marking; S13, construction of an NDVI time sequence data set; S2, space interpolation reconstruction S21, construction of a space neighborhood window; S22, calculation of an NDVI variation coefficient and an FMC variation coefficient in the space neighborhood, and screening of effective data; S23, inverse distance weighted interpolation; S24, generation of a space interpolation reconstruction result; S3, time interpolation reconstruction S31, time sequence FMC data extraction marking; S32, determination of a time interpolation window; S33, FMC data estimation on invalid data meaningful to time interpolation; S34, generation of a time interpolation result; and S4, space-time collaborative reconstruction. The application fully utilizes the correlation of FMC data in space and time, and realizes simple and efficient reconstruction.

Description

Technical Field

[0001] This application relates to the field of remote sensing drought monitoring and forest fire risk assessment technology, specifically a 500-meter spatial resolution fuel moisture content (FMC) prediction method based on spatial and temporal interpolation. This method uses global 500-meter resolution FMC data as a foundation, employing inverse distance weighting and exponential smoothing algorithms to reconstruct missing values ​​in the spatial and temporal dimensions, respectively. The reconstruction effects of different methods were systematically evaluated, and the effectiveness of spatiotemporal collaborative reconstruction was verified by comprehensively comparing the performance of spatial, temporal, and spatiotemporally weighted reconstruction strategies. Simultaneously, the improvement in data integrity before and after reconstruction was assessed, quantifying the improvement effect on the quality of the original data. Finally, more continuous global FMC data on an annual and daily scale was generated, providing higher-quality data support for forest fire risk early warning and drought prediction research. Background Technology

[0002] Fuel Moisture Content (FMC) is a key biophysical parameter characterizing the moisture state and flammability of vegetation, playing an important role in forest fire risk early warning, ecological drought monitoring, and terrestrial ecosystem process research. FMC reflects the response of vegetation to changes in meteorological conditions and serves as an important bridge connecting meteorological drought, soil drought, and vegetation drought.

[0003] With the development of remote sensing technology, satellite-based fire hazard (FMC) data has achieved global-scale coverage. For example, FMC data constructed based on MODIS data provides a data foundation for large-scale studies in terms of spatial resolution (approximately 500 meters) and temporal resolution (daily or composite periodicity). However, due to factors such as cloud cover, atmospheric conditions, sensor observation angle, and data quality control, existing FMC data generally suffer from varying degrees of data loss and outlier problems in both time series and spatial distribution. Specifically, this manifests as: ① discontinuous time series: some areas have long periods of missing data, making it difficult to support continuous dynamic analysis; ② fragmented spatial distribution: a large number of invalid pixels exist in single-moment images, affecting the representation of spatial structure; ③ high data noise: outliers and abrupt changes affect the analysis of FMC change trends; ④ insufficient data integrity: limiting its application in drought propagation and fire risk prediction.

[0004] To address the above issues, existing research typically employs interpolation methods to repair missing data, mainly including: (1) Spatial interpolation methods: such as inverse distance weighting (IDW) and Kriging, which estimate based on spatial neighborhood information, but these methods do not consider the characteristics of time variation and have limited effectiveness in the case of continuous time missing data; (2) Time interpolation methods: such as linear interpolation, spline interpolation and time series smoothing methods (such as Savitzky-Golay filtering, exponential smoothing, etc.), can restore the trend of time change, but it is difficult to maintain the consistency of spatial structure in areas with severe spatial defects; (3) Simple spatiotemporal interpolation method: Some studies attempt to combine time and space information, but most of them use fixed weights or simple superposition methods, lacking a systematic collaborative reconstruction mechanism, and it is difficult to maintain stable performance under different missing scenarios. Summary of the Invention

[0005] The purpose of this application is to address the problems of temporal discontinuity, severe spatial missingness, and unstable data quality in existing FMC data, and to propose an FMC data reconstruction method based on spatiotemporal interpolation. By co-modeling in both time and space dimensions, the method can achieve a continuous, complete, and stable representation of FMC data.

[0006] This application uses global 500-meter resolution FMC data as the benchmark data source and constructs a reconstruction method based on spatiotemporal interpolation to address the missing value problem in the temporal and spatial dimensions. This application does not rely on high-resolution remote sensing data or deep learning models, but directly uses the original FMC time series data to achieve continuous data reconstruction by mining its inherent spatiotemporal correlations. In the data processing, firstly, quality control and missing value identification are performed on the FMC data, and a pixel-by-pixel time series is constructed. For different types of missing data, interpolation models are established in both the spatial and temporal dimensions, and then further fused to obtain more stable and reliable reconstruction results.

[0007] The main principles of this application are as follows: This application leverages the significant continuity and correlation of Free Matrix Conversion (FMC) across the spatiotemporal dimensions to achieve reasonable estimation and reconstruction of missing data by constructing interpolation relationships in both spatial and temporal dimensions. Spatially, neighboring pixels typically exhibit strong spatial autocorrelation due to similarities in climate conditions, vegetation type, and topography. Therefore, when a pixel is missing at a specific time, its FMC values ​​can be estimated using distance-weighted methods, thus restoring the continuity of spatial distribution. Temporally, FMC typically exhibits a smooth dynamic evolution with seasonal changes, and its time series demonstrates strong autocorrelation and trends. Therefore, for missing values ​​in the time series, prediction can be made based on historical observation data using time-weighted methods, thereby restoring temporal continuity.

[0008] Based on the above principles, this application constructs interpolation models in the spatial and temporal dimensions respectively, and realizes spatiotemporal collaborative reconstruction through a fusion strategy, so that the reconstruction results simultaneously meet the requirements of spatial consistency and temporal continuity.

[0009] The vegetation combustible moisture content prediction method based on spatiotemporal interpolation described in this application mainly includes the following steps: S1: Data Acquisition and Preprocessing S11: Acquisition of Vegetation FMC Data Acquire vegetation FMC remote sensing data within the study period (in special cases, the unit is days) in the study area, and construct a daily time series FMC dataset; S12: Quality Control and Missing Data Identification (Marking Invalid Data) Based on the quality control bands in FMC data, low-quality and missing data are identified and marked as invalid data, and a time-series FMC dataset containing valid and invalid data is constructed. S13: Auxiliary Data Acquisition Obtain NDVI data that are spatially, temporally, and at the same resolution as the time-series FMC dataset, remove outliers, and construct the NDVI time-series dataset for subsequent spatial homogeneity constraints and parameter optimization.

[0010] S2: Spatial Interpolation Reconstruction S21: Spatial Neighborhood Construction For each invalid data point, a neighborhood window is constructed at its spatial location. The default window is 3×3, but in the verification experiment, it is expanded to 5×5 and 7×7 windows. Only valid FMC data is retained within the neighborhood. S22: Calculate the NDVI and FMC coefficients of variation within the spatial neighborhood, and filter valid data that meet the spatial homogeneity condition based on a preset threshold (valid data whose surrounding area meets the spatial homogeneity constraint can be used to determine valid data for spatial interpolation). Within a spatial neighborhood window, the NDVI-CV and FMC-CV coefficients are calculated, with dual threshold conditions set: NDVI-CV ≤ threshold 1 and FMC-CV ≤ threshold 2. Only when both threshold conditions are met simultaneously is the neighborhood considered a spatially homogeneous region, allowing interpolation. The formulas for calculating the two types of CVs (NDVI-CV and FMC-CV) are as follows: The coefficient of variation (CV) for a certain type of data is calculated as: CV = σ / μ × 100% Where σ represents the standard deviation of the overall data of a certain type of data (NDVI or FMC) within the spatial neighborhood window, and μ represents the mean of the overall data of that type of data (NDVI or FMC) within the spatial neighborhood window. That is, a certain coefficient of variation is the ratio of the standard deviation to the mean of that type of data within the spatial neighborhood window. S23: Inverse distance weighted interpolation Within a spatially homogeneous region of invalid data, inverse distance weighted interpolation is performed on the invalid data while simultaneously satisfying the following conditions: 1. Valid data from the spatially homogeneous region is used as input; 2. The weight of the valid data relative to the invalid data is calculated based on the spatial distance between the valid data used for interpolation and the invalid data; 3. The weight exponent is set to 2; 4. The weighted average of all valid data in the spatially homogeneous region is used as the spatial neighborhood reconstruction result of the invalid data; the inverse distance weighted interpolation formula is as follows:

[0011] Where, x p This represents the interpolated data at position p. Let d represent the i-th known valid data. i x represents p and The location distance is generally expressed as Euclidean distance, where β is the attenuation coefficient. This application typically involves performing step S21 on all invalid data for the entire research period in the study area, followed by step S22, and then step S23. Alternatively, steps S21 to S23 can be performed on one invalid data before performing steps S21 to S23 on another invalid data. The order in which these steps are performed does not affect the outcome.

[0012] S24: Generate and label the spatial interpolation results. Generate spatial interpolation to reconstruct FMC data and record the spatial neighborhood interpolation source identifiers; S3: Temporal Interpolation Reconstruction S31: Time-series FMC data extraction marker for invalid data For each invalid data point, extract its time-series FMC data and mark the missing time nodes; S32: Time window determination Based on different vegetation types and dates, the optimal time window before and after a certain date for the corresponding vegetation is dynamically determined (the optimal time window satisfies that the prediction accuracy R² is above the critical point or takes the longest time window with prediction accuracy R²>0.8). S33: Exponential smoothing interpolation For invalid data with meaningful temporal interpolation, a simple exponential smoothing method (SES) based on a sliding window is used to perform interpolation repair based on temporal adjacency. This application introduces a smoothing coefficient to make the weight of historical observations decrease exponentially with increasing time intervals, thus placing greater emphasis on the trend information contained in recent data. The formula for the simple exponential smoothing method is as follows:

[0013] Where, x t x represents the data to be interpolated at time t (invalid data). t+i σ represents the known valid data at time t+i within the current optimal time window for this vegetation. i x represents t+i With x t The time interval is m, where m represents the size of the one-way sliding window, and α is the smoothing coefficient.

[0014] The invalid data that is meaningful for time interpolation refers to: data that is in the corresponding vegetation growth season, has at least two valid data points within its optimal time window, and exists before and after the invalid data.

[0015] All invalid data in the study area of ​​this application should undergo the corresponding S31 to S33 steps throughout the entire study period. The order in which they are done does not matter. Alternatively, all invalid data can undergo the S31 step separately, then the S32 step separately, and finally the S33 step separately.

[0016] S34: Generate and label the time interpolation results. Generate time-series FMC reconstructed data with all invalid data and label the time interpolation sources; All invalid data in this application are spatially and / or temporally interpolated using only valid data. The FMC data obtained through spatial and / or temporal interpolation methods are not used for spatial and / or temporal interpolation reconstruction of other invalid data, but are only used for spatiotemporal interpolation reconstruction.

[0017] S4: Spatiotemporal Co-reconstruction and Labeling Temporal and spatial interpolation results are fused, reconstructed, and labeled. S41: When both spatial and temporal interpolation results of invalid data exist, the FMC (reconstructed) data of the invalid data is generated using a weighted average method and the source is marked. S42: When only one type of interpolation exists in the spatial and temporal interpolations of invalid data, the interpolation result is used as the FMC (reconstructed) data of the invalid data and its source is marked. For invalid data with both spatial and temporal interpolation results, a weighted average method is preferentially used for interpolation reconstruction. The weights of the weighted average method are initially calculated using the mathematical mean method. After obtaining the spatial and temporal interpolation accuracy for each vegetation species in step S53, the new weights are determined based on the comparison between the spatial and temporal interpolation accuracy of the vegetation species to which the invalid data belongs. That is, the spatial interpolation weight of vegetation species = spatial interpolation accuracy of vegetation species / (spatial interpolation accuracy of vegetation species + temporal interpolation accuracy of vegetation species), and the temporal interpolation weight of vegetation species = temporal interpolation accuracy of vegetation species / (spatial interpolation accuracy of vegetation species + temporal interpolation accuracy of vegetation species). When data in one dimension is completely missing, only the other dimension method is used to complete the data. No operation is performed when there are no spatial or temporal interpolations for invalid data, and the missing state is retained.

[0018] S5: Accuracy Assessment and Method Validation S51: Simulated Missing Experiment Randomly select 10% of the valid data from the daily time-series FMC dataset obtained in step S11 to create artificial invalid data (simulating missing data) through artificial masking, and construct a daily time-series FMC reconstruction test set containing at least 10% invalid data. S52: Based on the daily time series FMC reconstruction test set, three FMC data reconstruction and prediction experiments were conducted according to the time interpolation method, the spatial interpolation method, and the spatiotemporal weighted interpolation of this application (steps S2~S4); S53: Comparative Analysis of Reconstruction Results of Three Methods The accuracy of vegetation combustible water content obtained by three methods—temporal interpolation, spatial interpolation, and spatiotemporal weighted interpolation—was compared and analyzed using the calculated coefficient of determination (R²), standardized root mean square error (nRMSE), and standardized mean absolute error (nMAE). This included a comparison of the accuracy calculations for temporal interpolation, spatial interpolation, and spatiotemporal weighted interpolation for each vegetation species and for all vegetation species as a whole.

[0019] S6: FMC Reconstruction Data Acquisition and Application Analysis S611: Extract FMC reconstruction data from the subset of the study time period of the above-mentioned study area; S612: Using the comparison relationship between spatial interpolation accuracy and temporal interpolation accuracy obtained in step S53 as weights, repeat step S41 on the FMC reconstruction data that simultaneously has spatial interpolation data and temporal interpolation data, and use the recalculated FMC reconstruction data as the final FMC data to update the subset FMC reconstruction data in step S111. S62: Data Integrity Assessment The effect of data quality improvement is analyzed by calculating the original missing rate and the reconstruction filling rate. S63: Application Analysis of FMC Reconstruction Data from a Subset of the Study Period in the Study Area We analyzed all time-series FMC data and their changing trends of the reconstructed subset of the study area during the research period, and compared the differences before and after FMC reconstruction to verify the ecological rationality of FMC reconstruction.

[0020] Compared with existing technologies, this application has the following advantages and effects: it realizes the reconstruction of combustible moisture content (FMC) data based on spatiotemporal interpolation methods. By fully utilizing the temporal continuity and spatial correlation of FMC data, it effectively fills in the missing information in the original product, significantly improving the data integrity and temporal continuity, while maintaining the original spatial distribution characteristics and trends of FMC. This method does not rely on additional high-resolution data or complex models, has high computational efficiency, and a wide range of applications, providing continuous and reliable data support for global-scale forest fire risk early warning, drought monitoring, and ecological water process analysis. Attached Figure Description

[0021] Figure 1 This is the overall scheme for spatiotemporal filling of FMC data in this application. Figure 1 The left side shows a schematic diagram of spatial interpolation, where Sc and Sr represent the horizontal and vertical coordinates, respectively. Figure 1 The diagram on the right illustrates time interpolation, where the horizontal axis S represents spatial location and the vertical axis T represents time. Gray areas represent missing values, red areas represent missing values ​​to be filled, blue areas represent spatial sampled values, orange areas represent temporal sampled values, blank areas represent known data, blue borders represent spatial sampling areas, and orange borders represent temporal sampling areas.

[0022] Figure 2 This is a flowchart illustrating the interpolation method of this application.

[0023] Figure 3 To illustrate the changes in the proportion of NDVI data for different vegetation types within 3×3, 5×5, and 7×7 spatial neighborhood windows, reflecting the trend of data volume variation with the range of the coefficient of variation (CV) threshold, the values ​​on the horizontal axis corresponding to the solid points in the figure are set as the NDVI-CV threshold for the current vegetation type.

[0024] Figure 4 Scatter plots were generated for NDVI data predictions for different vegetation types within a 3×3 neighborhood window. The results show that the prediction accuracy varies among different vegetation types. 2 The values ​​range from 0.39 to 0.97, with most vegetation having R values. 2 The values ​​are concentrated in the high range of 0.7 to 0.8; using NDVI data as a preliminary verification and constraint tool for FMC data reconstruction not only effectively evaluates the applicability of the IDW method in vegetated areas, but also ensures the rigor of the interpolation process.

[0025] Figure 5 Scatter plots were generated for spatial prediction of FMC data for different vegetation types. To ensure the homogeneity of the interpolation source area, the spatial neighborhood window was strictly limited to 3×3. To overcome the potential shortcomings of relying solely on NDVI-CV constraints, the coefficient of variation (FMC-CV) of the FMC data itself was introduced as an auxiliary discrimination index. By calculating the FMC-CV of each pixel within the 3×3 window and statistically analyzing the global distribution, the FMC-CV value corresponding to 75% of the FMC data was selected as the second threshold. That is, when a pixel simultaneously meets the dual threshold conditions of NDVI-CV and FMC-CV, it is considered that the data fluctuation of the current area is low and can participate in FMC data spatial imputation. Within this optimized spatial constraint range, the missing value simulation and prediction process, consistent with the aforementioned NDVI verification stage, was executed. The predicted scatter plot density of FMC data for different vegetation types is shown below. Figure 5 As shown, the deciduous coniferous forest had the worst fitting effect (R0). 2 =0.4244), with open shrublands showing the highest accuracy (R = 0.4244). 2 =0.8510), with farmland / natural vegetation having the highest mean absolute error of 47 (%), while tropical grassland had the lowest error of 12 (%).

[0026] Figure 6 Taking Zhejiang Province in May 2016 as an example, this paper presents the trend of FMC prediction accuracy for various vegetation types as the time interval (number of days) changes, providing a basis for differentiated FMC prediction time parameters. The determination of the time interpolation interval in the application takes into account both prediction accuracy and interval length in a hierarchical decision-making process. The solid points correspond to the time intervals before and after the interpolation, which are used as the time interpolation parameters for each vegetation type in that month. Blue solid points are the first-choice critical points with the fastest rate of prediction accuracy decline. If the prediction accuracy at the current critical point is low, the next critical point with the second-lowest rate of decline is selected. For example, farmland / natural vegetation is selected as the second critical point, represented by green solid points; deciduous coniferous forests and deciduous broad-leaved forests are selected as the third critical points, represented by orange solid points. If none of the critical points meet the prediction accuracy requirements, the "quality first" principle is used, considering the longest time interval with a coefficient of determination greater than 0.8 on the test data as the interpolation parameter, such as evergreen broad-leaved forests and mixed forests, represented by red solid points in the figure. By selectively choosing sequence lengths and optimizing method parameters, the generated FMC time series data can be guaranteed to have higher robustness and reliability.

[0027] Figure 7The interpolation intervals before and after the optimal monthly values ​​for each vegetation type in Zhejiang Province in 2016 are shown. Most vegetation types exhibit stable short time windows for most months, with relatively stable FMC dynamics in the short term. Historical data and current data values ​​show high correlation, and a shorter interpolation window is sufficient to ensure accuracy. Longer optimal windows reveal the long-term stability of FMC changes when they are dominated by sustained and stable phenological or environmental trends. For example, during the hot summer months, the sustained high temperatures and water stress exert enormous pressure on the physiological activities of evergreen broad-leaved forests, leading to a slow and stable trend in FMC changes. Therefore, longer historical data can effectively improve the robustness of the prediction model.

[0028] Figure 8 This image compares the temporal, spatial, and spatiotemporal weighted prediction accuracy of different vegetation types in Zhejiang Province in August 2016. An image from August 15, 2016, was selected as the validation image. 10% of pixels for each vegetation type were randomly selected and masked. Spatial, temporal, and spatiotemporal weighted methods were used to reconstruct the missing data. R², nMAE, and nRMSE were used to evaluate the accuracy.

[0029] Figure 9 To showcase the regional effects of FMC data reconstruction in Zhejiang Province in 2016, FMC images of different regions taken on the 15th of each month in March, June, September, and December 2016 were selected for comparison.

[0030] Figure 10 This study aims to statistically analyze the proportion of missing and filled rates in daily FMC data of Zhejiang Province in 2016. Taking daily FMC data of Zhejiang Province in 2016 as an example, the study analyzes the original missing rate and the reconstructed filled rate.

[0031] Figure 11 This is a curve showing the daily average value of FMC data before and after reconstruction in China in 2016. Based on the global vegetation classification system of IGBP data, and taking China in 2016 as an example, a statistical analysis of the daily average value of FMC data is conducted. Detailed Implementation

[0032] The high spatial resolution vegetation combustible moisture content prediction method of this application mainly includes the following steps: S1: Data Acquisition and Preprocessing S11: FMC Data Acquisition Acquire FMC remote sensing data with a spatial resolution of 500 meters for a certain period of time in the study area (global or a specific region). In a special case, use the GEE platform to construct a daily time series FMC dataset using global time range (study period) FMC remote sensing data from 2003 to 2023. S12: Quality Control and Defect Detection Based on the quality control band (Quality_Level) in FMC data, low-quality pixels (data) and missing pixels (data) are identified and marked as invalid data, and an FMC time series dataset containing valid data and invalid data is constructed. S13: Auxiliary Data Acquisition Acquire NDVI data at a resolution of 500 meters from the same period, unify the spatial resolution with the FMC data, remove outliers (NDVI <-1 or >1), and construct NDVI time series data for subsequent spatial homogeneity constraints and parameter optimization.

[0033] S2: Spatial Interpolation Reconstruction S21: Spatial Neighborhood Construction For each invalid data point, a neighborhood window is constructed at its spatial location. The default window is 3×3, but in the validation experiment, it is expanded to 5×5 and 7×7 windows (7×7 window during sampling). Only valid FMC data is retained within the neighborhood. S22: Calculate the NDVI and FMC coefficients of variation within the spatial neighborhood, and filter valid data that meet the spatial homogeneity condition based on a preset threshold (valid data whose surrounding area meets the spatial homogeneity constraint can be used to determine valid data for spatial interpolation). To avoid introducing errors from heterogeneous regions, the NDVI coefficient of variation (NDVI-CV) and the FMC coefficient of variation (FMC-CV) are calculated within a spatial neighborhood window, with dual threshold conditions set: NDVI-CV ≤ threshold 1 and FMC-CV ≤ threshold 2. Only when both conditions are met simultaneously is the neighborhood considered a spatially homogeneous region, allowing interpolation. The CV calculation formula is as follows: CV = σ / μ × 100% Where σ represents the standard deviation of the overall data within the spatial neighborhood window, and μ represents the mean of the overall data within the spatial neighborhood window; S23: Inverse Distance Weighted Interpolation (IDW) Within a spatial neighborhood that satisfies the condition of spatial homogeneity, IDW interpolation is performed on invalid data: ① Use valid data from the spatial neighborhood as input; ② Calculate weights based on spatial distance; ③ Set the weight exponent to 2; ④ Calculate the weighted average as the spatial neighborhood reconstruction result; The basic formula of the IDW method is expressed as follows:

[0035] Where, x p This represents the interpolated data at position p. Let d represent the i-th known valid data. i x represents p and The location distance is generally expressed as Euclidean distance, where β is the attenuation coefficient. This application typically involves performing step S21 on all invalid data for the entire research period in the study area, followed by step S22, and then step S23. Alternatively, invalid data can be processed one by one using steps S21 to S23, and the order in which they are performed does not matter.

[0036] S24: Generate and label the spatial interpolation results. Generate time-series spatial interpolation reconstructed FMC data containing all invalid data, and record the spatial neighborhood interpolation source identifier (marking which valid data within the spatial neighborhood the spatial interpolation reconstructed FMC data comes from). S3: Temporal Interpolation Reconstruction S31: Time-series FMC data extraction marker for invalid data For each invalid data point, extract its time-series FMC data and mark the missing time nodes; S32: Time window determination Based on different vegetation types and dates, the optimal time window (referring to the time period, in days, around a certain date that best reflects the vegetation characteristics of that date) is dynamically determined: ① Analyze the changes in prediction accuracy under different time intervals to obtain the optimal time window (that day ± m days, where m is explained in step S33); ② Select the time interval corresponding to the critical point of the rate of decrease in prediction accuracy (coefficient of determination R²); ③ If the accuracy is insufficient, adopt the longest time window that satisfies R²>0.8; ④ Adopt a differentiated time window strategy for different vegetation types; ②, ③, and ④ are detailed in [link to relevant documentation]. Figure 6 Explanation: Taking May as an example, ②: solid blue dots As the primary critical point with the fastest rate of decline in prediction accuracy, if the prediction accuracy is low at the time intervals before and after the current critical point, then the critical point with the second fastest rate of decline is selected. For example, farmland / natural vegetation is selected as the second critical point. Green Reality Heart Points Indicated; deciduous coniferous forests and deciduous broad-leaved forests were selected as the third critical point, using Orange solid dots ③: If all critical points fail to meet the prediction accuracy requirements, the "quality first" principle is used, considering the longest time interval before and after the test data with a coefficient of determination greater than 0.8 as the interpolation parameter, such as evergreen broad-leaved forests and mixed forests. This is indicated in the graph using... solid red dot This indicates that by differentially selecting sequence lengths and optimizing method parameters, the generated FMC time series data can be guaranteed to have higher robustness and reliability. The table below shows... Figure 6 Detailed data on the precision changes of each vegetation type at different time intervals (where the number of types corresponds to...) Figure 6 The number in parentheses, and d in the first line represents the time interval in days.

[0037] Type 1 0.9339 0.8974 0.8553 0.7880 0.7474 0.6980 0.6272 0.5617 0.5052 0.4298 0.3442 0.2284 0.1196 0.0291 0.0291 Type 2 0.9616 0.9508 0.9410 0.9004 0.9015 0.8856 0.8723 0.8597 0.8348 0.8187 0.7984 0.7902 0.7813 0.7658 0.7480 Type 3 0.8687 0.7880 0.6947 0.4273 0.3084 0 0 0 0 0 0 0 0 0 0 Type 4 0.9517 0.9401 0.9054 0.8472 0.8196 0.8050 0.7581 0.7265 0.6661 0.6142 0.5510 0.4920 0.4235 0.3971 0.3510 Type 5 0.8837 0.8249 0.8014 0.7324 0.6673 0.6415 0.5695 0.4350 0.3537 0.1775 0.1775 0.1775 0.1775 0.1775 0.1775 Type 6 0.8067 0.7638 0.7260 0.6206 0.6777 0.6495 0.6269 0.5960 0.5502 0.5168 0.4975 0.4774 0.4766 0.4570 0.4096 Type 7 0.9547 0.9335 0.9125 0.8796 0.8615 0.8429 0.8253 0.8057 0.7861 0.7664 0.7489 0.7305 0.7136 0.6982 0.6814 Type 8 0.9480 0.9273 0.9031 0.8477 0.8251 0.7992 0.7603 0.7214 0.6620 0.5982 0.4978 0.3711 0.2940 0.1197 0.0872 Type 9 0.9477 0.9290 0.9057 0.8584 0.8403 0.8185 0.7962 0.7730 0.7386 0.7067 0.6646 0.6240 0.5956 0.5574 0.5239 Type 10 0.9648 0.9509 0.9335 0.8996 0.8863 0.8532 0.8375 0.8180 0.8180 0.7965 0.7755 0.7546 0.7408 0.7227 0.7085 Type 12 0.9571 0.9399 0.9217 0.8712 0.8608 0.8431 0.8242 0.8062 0.7842 0.7617 0.7420 0.7157 0.6991 0.6737 0.6491 Type 14 0.9605 0.9468 0.9292 0.8677 0.8530 0.8310 0.8047 0.7751 0.7247 0.6751 0.6139 0.5455 0.5229 0.4772 0.4243 S33: Exponential Smoothing Interpolation (SES) For invalid data that is meaningful for time interpolation, a simple exponential smoothing method (SES) based on a sliding window is used to interpolate and repair the invalid data. This application is based on the assumption that "time-adjacent points are more correlated." The SES method introduces a smoothing coefficient, which causes the weight of historical observations to decrease exponentially with the increase of time interval, placing more emphasis on the trend information contained in recent data. The basic formula is expressed as follows:

[0038] Where, x t x represents the data to be interpolated at time t (invalid data). t+i σ represents the known valid data at time t+i within the current optimal time window for this vegetation. i x represents t+i With x t The time interval is m, which represents the size of the one-way sliding window, and α is the smoothing coefficient, which is 0.6.

[0039] S34: Generate and label the time interpolation results. Generate time-series FMC reconstructed data with all invalid data and label the time interpolation sources; S4: Spatiotemporal Co-reconstruction and Labeling Temporal and spatial interpolation results are fused, reconstructed, and labeled. S41: When both spatial and temporal interpolation results of invalid data exist, the FMC (reconstructed) data of the invalid data is generated using a weighted average method and the source is marked. S42: When only one type of interpolation exists in the spatial and temporal interpolations of invalid data, the interpolation result is used as the FMC (reconstructed) data of the invalid data and its source is marked. For invalid data with both spatial and temporal interpolation results, a weighted average method is preferentially used for interpolation reconstruction. The weights of the weighted average method are initially calculated using the mathematical mean method. After obtaining the spatial and temporal interpolation accuracy for each vegetation variety in step S53, the new weights are determined based on the comparison between the spatial and temporal interpolation accuracy of the invalid data variety. That is, the spatial interpolation weight of any vegetation variety = spatial interpolation accuracy of that vegetation variety / (spatial interpolation accuracy of that vegetation variety + temporal interpolation accuracy of that vegetation variety), and the temporal interpolation weight of any vegetation variety = temporal interpolation accuracy of that vegetation variety / (spatial interpolation accuracy of that vegetation variety + temporal interpolation accuracy of that vegetation variety). When data in one dimension is completely missing, only the other dimension method is used to complete it. When neither dimension of data is available, no spatiotemporal interpolation is performed.

[0040] S5: Accuracy Assessment and Method Validation S51: Simulated Missing Experiment Randomly select 10% of the valid data in the daily time series FMC dataset obtained in step S11 to create artificial invalid data through artificial masking, and construct a daily time series FMC reconstruction test set containing at least 10% invalid data; S52: Based on the daily time series FMC reconstruction test set, three FMC data reconstruction and prediction experiments were conducted according to the time interpolation method, the spatial interpolation method, and the spatiotemporal weighted interpolation of this application (steps S2~S4); S53: Comparative Analysis of Reconstruction Results of Three Methods The accuracy of vegetation combustible water content obtained by the three methods of time interpolation, spatial interpolation, and spatiotemporal weighted interpolation was compared (including the accuracy calculation and comparison of time interpolation, spatial interpolation, and spatiotemporal weighted interpolation for each vegetation species and for all vegetation as a whole) was evaluated.

[0041] S6: FMC Reconstruction Data Acquisition and Application Analysis S61: FMC reconstruction data obtained from a subset of the study period in the study area S611: Extract FMC reconstruction data (all FMC data after spatiotemporal interpolation) from the subset of the study time period of the above study area. S612: Using the comparison relationship between spatial interpolation accuracy and temporal interpolation accuracy obtained in step S53 as weights, repeat step S41 on the FMC reconstruction data that has both spatial and temporal interpolation data. Use the recalculated FMC reconstruction data as the final FMC data and update the subset of FMC reconstruction data in step S111 (only update the FMC reconstruction data that has both spatial and temporal interpolation data).

[0042] S62: Data Integrity Assessment The data quality improvement effect is analyzed by calculating the original missing rate and the reconstruction filling rate. The original missing rate is defined as the ratio of the number of invalid data in all FMC data to the total amount of FMC data, reflecting the data integrity problem of the original data. The reconstruction filling rate is defined as the ratio of the number of reconstructed FMC data to the number of original invalid data, reflecting the reconstruction capability of invalid data. Figure 10 This reflects the original missing data rate and the reconstruction and filling rate. The red curve is the reconstruction and filling rate curve. It can be seen that the reconstruction and filling rate varies in different months, with the highest reconstruction and filling rate reaching 50% during the growing season in Zhejiang Province.

[0043] S63: Application Analysis of the Study Area We selected a reconstructed subset (such as Zhejiang Province and China) within the study period of the study area, analyzed its time series FMC change trend, and compared the differences before and after FMC reconstruction to verify the ecological rationality of FMC reconstruction.

[0044] The invalid data that is meaningful for time interpolation as described in this application refers to: data that is in the corresponding vegetation growing season and has at least two valid data points within its optimal time window, and that exists before or after the invalid data (not just before or after the invalid data); cases where the time series FMC information data is severely missing and / or not distributed before or after the invalid data are excluded from the non-growing season and / or within the optimal time window. These cases retain their original missing state to avoid low-quality imputation.

[0045] The following analysis of the results further describes this application: An image from August 15, 2016, was selected as the verification image. 10% of the pixels for each vegetation type were randomly selected and masked. Spatial, temporal, and spatiotemporal weighted methods were used to reconstruct the missing data. R², nMAE, and nRMSE were used to evaluate the accuracy. The results are presented in the form of a heatmap. Figure 8 This study provides a direct comparison of the prediction differences between different methods and different vegetation types. The results show that in areas with high spatial heterogeneity, such as vegetation ecotones or fragmented terrain, spatial prediction performance is low. For example, the R² value for the farmland / natural vegetation patchwork type is 0.729, and the error index is relatively high. For vegetation types with clear phenological patterns, especially within specific phenological stages, temporal interpolation shows a more significant advantage; for example, the monthly prediction R² values ​​for grassland and cultivated land types all reach above 0.9. Considering all three evaluation indicators, the spatiotemporal weighted method performs best for most vegetation types, with its R² value being closest to 1 in most cases, and its nMAE and nRMSE values ​​generally lower than those of the single-dimensional method. Based on these conclusions, to ensure reconstruction accuracy while maximizing the use of existing data resources, the spatiotemporal weighted method should be prioritized for interpolation reconstruction when data conditions permit; when data in one dimension is completely missing, only the other dimension method should be used to complete the data.

[0046] This application selects FMC images from different regions on the 15th of each month in March, June, September, and December 2016 for front-to-back comparison. Figure 9The results show that after reconstruction, the spatial coverage of the FMC data was significantly improved, effectively filling data gaps in the original imagery caused by cloud cover and missing observations. Simultaneously, the reconstructed FMC exhibits good continuity and consistency in its spatial distribution pattern, clearly reflecting spatial differences in vegetation moisture conditions across different regions. From an ecological perspective, the reconstruction results show variation characteristics consistent with vegetation growth patterns in different seasons. For example, in March and June, the overall FMC data level is high, reflecting vigorous vegetation growth and favorable moisture conditions; while in September and December, the FMC data in some areas declines, reflecting seasonal drought or vegetation decline.

[0047] This application uses daily FMC data from Zhejiang Province in 2016 as an example to analyze the original missing rate and reconstruction filling rate. The original missing rate is defined as the ratio of invalid pixels in the daily FMC data to the total number of vegetation pixels in the study area, reflecting the data integrity issues of the original data. Figure 10 The study found that the original missing data rate curve exhibited significant seasonal variations, generally showing a "V"-shaped trend, with higher missing data rates in spring and winter and significantly lower rates in summer and autumn. The filling rate curve showed a negative correlation with the missing data rate curve; that is, the higher the missing data rate, the lower the filling rate. During periods of severe continuous data loss, the limited spatiotemporal information in the filling process restricted the method's repair potential. The results of this application confirm that the spatiotemporal reconstruction method, while ensuring the reliability of the results, achieves a maximum filling rate of over 50%, meeting the requirement of significantly improving the FMC missing data problem.

[0048] This application takes the China region in 2016 as an example to conduct statistical analysis on the daily average FMC (Frequency Control Metric). Figure 11 The figures show the daily average variation curves of the original, reconstructed, and combined data for broadleaf forests, coniferous forests, shrubs, and grasslands, respectively, exhibiting an overall dynamic characteristic of "low in winter, rising in spring, high in summer, and falling in autumn." The reconstructed data curves for different vegetation types show a high degree of consistency with the original data curves in terms of variation trends, but there are certain differences in numerical magnitude and fluctuation details, reflecting the spatiotemporal reconstruction method's responsiveness to local environmental factors and its sensitive capture of short-term processes.

[0049] In summary, this application employs the IDW interpolation algorithm in the spatial dimension, introducing the coefficient of variation of NDVI data and the coefficient of variation of FMC data itself to construct a dual spatial constraint threshold. In the temporal dimension, it adopts the SES method, proposing a differentiated hierarchical strategy for determining the optimal temporal interpolation window based on the phenological patterns of different vegetation types. Through controlled experiments based on sampled data and application analysis based on real remote sensing data, this application comprehensively verifies the effectiveness of the spatiotemporal reconstruction method, increasing the maximum fill rate of the original data to over 50%, successfully generating more continuous and reliable FMC spatiotemporal datasets, with a maximum data fill rate exceeding 50%, providing a data foundation for conducting large-scale dynamic fire risk assessment and other related research.

Claims

1. A method for predicting the moisture content of vegetation combustibles based on spatiotemporal interpolation, characterized in that... Includes the following steps: S1: Data Acquisition and Preprocessing S11: Acquisition of Vegetation FMC Data Acquire vegetation FMC remote sensing data within the study period of the study area and construct a daily time series FMC dataset; S12: Mark invalid data Based on the quality control bands in FMC data, low-quality and missing data are identified and marked as invalid data, and a time-series FMC dataset containing valid and invalid data is constructed. S13: Auxiliary Data Acquisition Obtain NDVI data that are spatially, temporally, and at the same resolution as the time-series FMC dataset, and remove outliers to construct the NDVI time-series dataset. S2: Spatial Interpolation Reconstruction S21: For invalid data, construct a spatial neighborhood window centered on the invalid data; S22: Calculate the NDVI and FMC coefficients of variation within the spatial neighborhood, and filter valid data that meet the spatial homogeneity condition based on a preset threshold. S23: Within the spatial neighborhood window of invalid data, perform inverse distance weighted interpolation on the invalid data using valid data that satisfies the spatial homogeneity condition; S24: Generate and label the spatial interpolation results; S3: Temporal Interpolation Reconstruction S31: Time series FMC data extraction marker for invalid data; S32: Determine the temporal interpolation window for invalid data based on vegetation type and time scale characteristics; S33: Use a simple exponential smoothing method to estimate FMC data for invalid data that are meaningful for time interpolation; S34: Generate and label the time interpolation results; S4: Spatiotemporal Co-construction.

2. The method for predicting the moisture content of vegetation combustibles according to claim 1, characterized in that: The spatial homogeneity determination rule mentioned in step S22 is: Within a spatial neighborhood window, calculate the NDVI coefficient of variation (NDVI-CV) and the FMC coefficient of variation (FMC-CV), and set dual threshold conditions: NDVI-CV ≤ threshold 1 and FMC-CV ≤ threshold 2. Only when the dual threshold conditions are met simultaneously is the neighborhood determined to be a spatially homogeneous region, allowing interpolation. The coefficient of variation of a certain type of data is the ratio of the standard deviation to the mean of that type of data within the spatial neighborhood window.

3. The method for predicting the moisture content of vegetation combustibles according to claim 1, characterized in that: The inverse distance weighted interpolation method described in step S23 is as follows: Within a spatially homogeneous region of invalid data, inverse distance weighted interpolation is performed on the invalid data while simultaneously satisfying the following conditions:

1. Valid data from the spatially homogeneous region is used as input; 2. The spatial distance between the valid data used for interpolation and the invalid data is used as the weight of the valid data relative to the invalid data; 3. The weight exponent is set to 2; 4. The weighted average of all valid data in the spatially homogeneous region is used as the spatial neighborhood reconstruction result of the invalid data. The inverse distance weighted interpolation formula is as follows:

4. Among them, x p This represents the interpolated data at position p. Let d represent the i-th known valid data. i x represents p and The Euclidean distance, where β is the attenuation coefficient.

5. The method for predicting the moisture content of vegetation combustibles according to claim 1, characterized in that: The formula for the simple exponential smoothing method described in step S33 is as follows:

6. Among them, x t x represents the data to be interpolated at time t. t+i σ represents the known valid data at time t+i within the current optimal time window for this vegetation. i x represents t+i With x t The time interval is m, where m represents the size of the one-way sliding window, and α is the smoothing coefficient.

7. The method for predicting the moisture content of vegetation combustibles according to claim 1, characterized in that: In step S33, invalid data that is meaningful for time interpolation refers to data that is in the corresponding vegetation growth season, has at least two valid data within its optimal time window, and exists before and after the invalid data.

8. The method for predicting the moisture content of vegetation combustibles according to claim 1, characterized in that: Step S4 includes: S41: When both spatial and temporal interpolation results exist, FMC data is generated using a weighted average method; S42: When only one type of interpolation exists between spatial interpolation and temporal interpolation, the interpolation result is used as the FMC data.

9. The method for predicting the moisture content of vegetation combustibles according to claim 4, characterized in that: The smoothing coefficient is 0.

6.

10. The method for predicting the moisture content of vegetation combustibles according to claim 1, characterized in that: It also includes an S5 step: S5: Accuracy assessment and method validation, including S51: Construct a daily time-series FMC reconstruction test set using a random mask method; S52: Based on the daily time series FMC reconstruction test set, three FMC data reconstruction and prediction experiments were conducted using time interpolation, spatial interpolation, and spatiotemporal weighted interpolation methods respectively. S53: Comparative analysis of the reconstruction effects of the three methods and comparison of the accuracy calculations of spatial interpolation and temporal interpolation. The accuracy of vegetation combustible water content obtained by three methods—time interpolation, spatial interpolation, and spatiotemporal weighted interpolation—was calculated and compared using the coefficient of determination, standardized root mean square error, and standardized mean absolute error as standards.

11. The method for predicting the moisture content of vegetation combustibles according to claim 8, characterized in that: It also includes an S6 step: S6: FMC Reconstruction Data Acquisition and Application Analysis, including S61: FMC reconstruction data acquisition; S62: FMC Reconstruction Data Integrity Assessment.

12. The method for predicting the moisture content of vegetation combustibles according to claim 9, characterized in that: Step S61 includes: S611: Extract FMC reconstruction data from a subset of the study time period in the study area; S612: Using the comparison relationship between the spatial interpolation accuracy and temporal interpolation accuracy of vegetation varieties obtained in step S53 as weights, the FMC reconstruction data that simultaneously possesses spatial interpolation data and temporal interpolation data are re-performed in step S41 according to the new weights. The recalculated FMC reconstruction data is used as the final FMC data, and the FMC reconstruction data corresponding to the study time period subset of the study area in step S611 is updated. The new weights refer to the following: vegetation variety spatial interpolation weight = vegetation variety spatial interpolation accuracy / (vegetation variety spatial interpolation accuracy + vegetation variety temporal interpolation accuracy), and vegetation variety temporal interpolation weight = vegetation variety temporal interpolation accuracy / (vegetation variety spatial interpolation accuracy + vegetation variety temporal interpolation accuracy).

Images