Method and device for detecting fast-growing forest felling taking into account sentinel-1 temporal autocorrelation

Abstract

The present application aims at the problem that the Sentinel-1 backscattering data is significantly short-term time autocorrelated due to the influence of rainfall, soil moisture and surface water content change, and further leads to the high false alarm rate of traditional difference cumulative sum and deforestation detection method, discloses a kind of fast-growing forest deforestation detection method and device considering Sentinel-1 time correlation, and proposes an improved difference cumulative sum detection method by fusing sliding time window and random block resampling significance test, which calculates local difference cumulative sum curve using sliding time window, and extracts candidate change points based on maximum value;To overcome the statistical test bias caused by time correlation, a random block resampling strategy is innovatively introduced to retain the local dependence structure of time series, and an extreme difference null hypothesis distribution is constructed for significance test.The present application effectively improves the precision and robustness of short-period fast-growing forest deforestation event automatic identification in large-scale, long-time series monitoring scenarios.

Description

Technical Field

[0001] This invention relates to the field of remote sensing time series change detection technology, and more specifically, to a method and apparatus for detecting the felling of fast-growing forests that takes into account the time series autocorrelation of Sentinel-1. Background Technology

[0002] Within the framework of precise forest resource management and digital supervision, capturing the specific spatiotemporal displacement of forest stand harvesting is a crucial logical starting point for forestry management decisions. Especially in tropical and subtropical regions, short-cycle, fast-growing plantations, represented by eucalyptus, exhibit extremely high management activity: their rotation cycles are extremely short and their generational spatial distribution is highly fragmented. This "small-scale, high-frequency, and asynchronous" harvesting characteristic results in significant randomness and suddenness in forest cover changes at the landscape scale. Traditional monitoring methods, due to insufficient spatiotemporal resolution or cloud cover, often exhibit severe signal lag when dealing with such dynamic changes, making it difficult to support time-sensitive business scenarios such as forest land asset verification, timber supply chain traceability, and illegal logging early warning. Sentinel-1 synthetic aperture radar, with its penetrating imaging and high-frequency revisit capabilities, can characterize stand structure loss by monitoring dramatic fluctuations in the backscattering coefficient, and has become a strategic data source for achieving near-real-time forest dynamic monitoring.

[0003] In long-sequence remote sensing dynamic monitoring, the differential cumulative sum algorithm, by characterizing the cumulative deviation features of the observed sequence over time, can effectively respond to sudden changes in the overall state of the time series, and is therefore widely used in the identification and location of structural abrupt events such as deforestation and degradation. However, when processing Sentinel-1 time series, the traditional differential cumulative sum algorithm faces the failure of statistical assumptions. First, there is the aliasing of random environmental disturbances: SAR signals not only respond to structural changes in biomass, but are also extremely sensitive to rainfall events, vegetation water content, and soil moisture. These instantaneous, non-stationary fluctuations induced by environmental factors often highly overlap with the structural response signals in the early stages of logging in the spectrum. Second, there is the interference of temporal autocorrelation: although high-frequency revisits improve temporal resolution, they also introduce significant observation dependence, directly undermining the core assumption of "independent and identically distributed" in statistical tests. In actual monitoring, this correlation causes the statistical distribution to drift, making it difficult for the algorithm to distinguish between real logging signals and background meteorological noise, resulting in a very high false alarm rate and severely weakening the production practicality of the monitoring system.

[0004] Existing optimization techniques often rely on temporal smoothing filters or heuristic empirical thresholds to suppress background fluctuations. These approaches are essentially empirically driven deterministic models, failing to systematically analyze the stochastic evolution and inherent autocorrelation structure of backscattering time series from a probabilistic modeling perspective. Particularly in the significance testing phase, most methods still employ theoretical inferences based on asymptotic normal distributions or naive random resampling methods. This approach completely ignores the temporal dependence of forest background signals in "unchanging" states, resulting in detectors lacking the necessary robustness in the face of complex climate fluctuations. Summary of the Invention

[0005] Based on the analysis of existing technologies, the key scientific problem that urgently needs to be solved in the current automated forestry remote sensing monitoring is how to construct a statistical inference scheme that can take into account both temporal correlation and non-normal distribution characteristics in large-scale, long-term monitoring, so as to achieve reliable extraction of real logging events.

[0006] Based on this, this invention proposes a method for detecting logging in fast-growing plantations that takes into account the time-series autocorrelation of Sentinel-1. This method uses pixel-level long-term backscatter data acquired by Sentinel-1 synthetic aperture radar as its foundation. By introducing a sliding time window mechanism in the time dimension, it locally models and segments the variation characteristics of the backscatter time series within different time intervals. Furthermore, it effectively enhances and characterizes the structural changes in the time series by combining cumulative difference and statistical measures, thereby improving the response capability to abrupt signals caused by logging. Simultaneously, in the process of determining the significance of changes, a random block resampling test mechanism that preserves the local dependency structure of the time series is adopted to test the statistical significance of candidate change points, addressing the statistical correlation commonly present between adjacent observations in the synthetic aperture radar data time series. This ensures the sensitivity of change detection while effectively suppressing spurious change responses caused by meteorological disturbances and natural fluctuations, thus achieving robust and automated identification of short-period fast-growing forest logging events.

[0007] To achieve the above objectives, a first aspect of the present invention provides a method for detecting logging in fast-growing forests that takes into account the temporal autocorrelation of Sentinel-1, comprising:

[0008] Acquire multi-temporal Sentinel-1 synthetic aperture radar images of the target area, construct a multi-year long-time series dataset with spatiotemporal consistency, and preprocess the constructed multi-year long-time series dataset;

[0009] Based on the preprocessed multi-year long-time series dataset, a time series of backscattering coefficients is constructed.

[0010] Based on the rotation cycle and growth characteristics of short-cycle fast-growing plantations, a sliding time window parameter was set to divide the backscattering coefficient time series into sliding segments.

[0011] Based on the partitioned sliding time window, construct the original difference sequence and the cumulative difference sequence;

[0012] The true range is calculated based on the constructed cumulative difference sum sequence, and the time point corresponding to the maximum value of the cumulative difference sum is extracted.

[0013] A continuous subsequence of a preset length is randomly selected from the original difference sequence, and a resampled difference sequence is constructed using sampling with replacement.

[0014] Calculate the cumulative sum of resampled differences based on the resampled difference sequence, and construct the set of range null hypothesis distributions;

[0015] Based on the relationship between the true range and the set of distributions with the null hypothesis of the range, a statistical significance test is performed on the candidate change points;

[0016] Based on the significance test results and the time interval corresponding to the maximum value of the cumulative sum of differences, determine whether the candidate change point is a felling point.

[0017] In one implementation, the constructed multi-year long-time-series dataset is preprocessed, including:

[0018] Standardize the image data in the multi-year long-time series dataset;

[0019] Distance-Doppler terrain correction using digital elevation models;

[0020] A multi-temporal speckle filtering algorithm is used to reduce the inherent speckle noise in synthetic aperture radar images;

[0021] Convert radar digital quantization values ​​into backscattering coefficients.

[0022] In one implementation, a backscattering coefficient time series is constructed based on a preprocessed multi-year long-term time series dataset, including:

[0023] Arrange the images in the preprocessed multi-year long-time series dataset in chronological order;

[0024] For each pixel within the target area, its backscattering coefficient values ​​are extracted across all effective time phases to construct a backscattering coefficient time series.

[0025] In one implementation, based on the divided sliding time window, the original difference sequence and the cumulative difference sequence are constructed, including:

[0026] The mean of all observations included in the divided sliding time window is calculated and used as the baseline reference value for the un-logged state within the window;

[0027] Calculate the deviation of each observation from the mean within the sliding time window, and construct the original difference sequence;

[0028] The original difference sequence is accumulated along the time sequence to obtain the cumulative difference sequence within the sliding time window. The curve shape of the cumulative difference sequence is used to reflect the structural change trend of the time series.

[0029] In one implementation, the true range is calculated based on the constructed cumulative difference sum sequence, and the time point corresponding to the maximum value of the cumulative difference sum is extracted, including:

[0030] Based on the constructed cumulative sum of differences sequence, the maximum and minimum values ​​in the cumulative sum sequence within the sliding time window are calculated. The true range is then calculated based on the maximum and minimum values ​​in the cumulative sum sequence, serving as the core feature for determining whether structural changes exist.

[0031] The time point at which the cumulative difference sequence reaches its maximum value is taken as the potential time point at which the logging event is most likely to occur within the sliding time window.

[0032] In one implementation, a continuous subsequence of a preset length is randomly selected from the original difference sequence, and a resampled difference sequence is constructed using sampling with replacement, including:

[0033] A random block resampling strategy is adopted, and the length of the resampling block is set based on the temporal correlation of the backscattering coefficient time series.

[0034] Using sampling with replacement, a continuous subsequence of a set length is randomly selected from the original difference sequence. The selected continuous subsequences are then spliced ​​together in the order of selection until the length of the spliced ​​new sequence reaches or exceeds the window length.

[0035] Extract the number of data points corresponding to the window length from the spliced ​​new sequence to generate a resampled difference sequence.

[0036] In one implementation, a cumulative sum of resampled differences is calculated based on the resampled difference sequence, and a set of range null hypothesis distributions is constructed, including:

[0037] For the generated resampled difference sequence, the resampled difference sequence is accumulated along the time sequence to obtain the resampled difference cumulative sum sequence;

[0038] The range is calculated based on the cumulative sum of resampled differences. The range obtained from multiple samplings and calculations constitutes the set of the null hypothesis distribution of the range.

[0039] In one implementation, based on the relationship between the true range and the set of distributions with the null hypothesis of the range, a statistical significance test is performed on the candidate change points, including:

[0040] The confidence level is calculated by comparing the true range calculated based on the original data with the set of distributions of the null hypothesis of the range.

[0041] If the confidence level is less than or equal to the set significance threshold, the result is considered statistically significant, indicating a significant change.

[0042] Based on the same inventive concept, a second aspect of the present invention provides a fast-growing forest deforestation detection device that takes into account the temporal autocorrelation of Sentinel-1, comprising:

[0043] The data acquisition and preprocessing module is used to acquire multi-temporal Sentinel-1 synthetic aperture radar images of the target area, construct a multi-year long-time series dataset with spatiotemporal consistency, and preprocess the constructed multi-year long-time series dataset.

[0044] The backscattering coefficient time series construction module is used to construct backscattering coefficient time series based on preprocessed multi-year long time series datasets.

[0045] The sliding time window division module is used to set sliding time window parameters based on the rotation cycle and growth characteristics of short-cycle fast-growing plantations, and to divide the backscattering coefficient time series successively according to the set sliding time window parameters.

[0046] The difference cumulative sum sequence construction module is used to construct the original difference sequence and the difference cumulative sum sequence based on the partitioned sliding time window;

[0047] The module for calculating the true range and extracting time points is used to calculate the true range based on the constructed cumulative difference sum sequence, which serves as the core feature for determining whether there is a structural change, and to extract the time point corresponding to the maximum value of the cumulative difference sum.

[0048] The resampled difference sequence construction module is used to randomly extract a continuous subsequence of a preset length from the original difference sequence and construct the resampled difference sequence using sampling with replacement.

[0049] The range null hypothesis distribution set construction module is used to calculate the cumulative sum of resampled differences based on the resampled difference sequence and construct the range null hypothesis distribution set;

[0050] The significance test module is used to perform statistical significance tests on candidate change points based on the relationship between the true range and the set of distributions of the null hypothesis of the range.

[0051] The logging time point location module is used to determine whether a candidate change point is a logging point based on the interval between the time points corresponding to the maximum value of the cumulative sum of differences in the significance test results.

[0052] Based on the same inventive concept, a third aspect of the present invention provides a computer device, including 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 fast-growing forest logging detection method considering Sentinel-1 time-series autocorrelation described in the first aspect.

[0053] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows:

[0054] This invention discloses a method for detecting logging in fast-growing plantations that takes into account the temporal autocorrelation of Sentinel-1 data. By constructing a spatiotemporally consistent long-term series dataset of Sentinel-1 backscattering coefficients, and utilizing a sliding time window combined with a cumulative difference algorithm, the method locally characterizes and accurately locates the structural change trends in the time series. Simultaneously, addressing the common statistical correlation between adjacent observations in synthetic aperture radar (SAR) data time series, an innovative random block resampling strategy based on temporally continuous sub-blocks is introduced. The cumulative difference statistic is then subjected to significance testing, effectively preserving the local dependency structure of the original time series during resampling. A null hypothesis distribution that more closely resembles the statistical characteristics of a real non-logging noise environment is constructed, effectively reducing the interference of short-term fluctuations caused by meteorological factors such as rainfall on the change detection results. Under complex meteorological conditions, it significantly suppresses false alarms caused by temporal autocorrelation, thereby improving the accuracy and robustness of automated identification of short-period fast-growing forest logging events. This method is suitable for large-scale, long-term dynamic monitoring and assessment of forest logging. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of a fast-growing forest felling detection method that takes into account the temporal autocorrelation of Sentinel 1 in an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of the detection results of eucalyptus plantation felling in one embodiment of the present invention;

[0058] Figure 3 This is a structural block diagram of a fast-growing forest logging detection device that takes into account the temporal autocorrelation of Sentinel No. 1 in an embodiment of the present invention. Detailed Implementation

[0059] This invention addresses the problem of high false alarm rates in traditional difference accumulation and deforestation detection methods due to significant short-term temporal autocorrelation in Sentinel-1 backscatter data caused by variations in rainfall, soil moisture, and surface water content. It proposes an improved difference accumulation detection method that integrates a sliding time window and random block resampling significance testing. This method uses a sliding time window to calculate local difference accumulation curves and extracts candidate change points based on the maximum value. To overcome statistical test bias caused by temporal autocorrelation, a novel random block resampling strategy that preserves the local temporal dependency structure is introduced to construct a range null hypothesis distribution for significance testing. This invention effectively improves the accuracy and robustness of automated identification of short-cycle fast-growing forest deforestation events in large-scale, long-term monitoring scenarios.

[0060] Example 1

[0061] This embodiment provides a method for detecting logging in fast-growing forests that takes into account the temporal autocorrelation of Sentinel-1. Please refer to [link to relevant documentation]. Figure 1 ,include:

[0062] S1: Acquire multi-temporal Sentinel-1 synthetic aperture radar images of the target area, construct a multi-year long-time series dataset with spatiotemporal consistency, and preprocess the constructed multi-year long-time series dataset.

[0063] Specifically, preprocessing of the constructed multi-year long-term time series dataset includes standardization, correction, and filtering.

[0064] In one implementation, S1 includes:

[0065] S1.1: Standardize the image data in the multi-year long-term dataset;

[0066] S1.2: Range-Doppler terrain correction using digital elevation model;

[0067] S1.3: Employ a multi-temporal speckle filtering algorithm to reduce the inherent speckle noise in synthetic aperture radar images;

[0068] S1.4: Convert radar digital quantization values ​​into backscattering coefficients.

[0069] Specifically, S1.1 standardizes the image data in the multi-year long-term dataset to ensure the continuity and consistency of the data in terms of geographical scope and time series.

[0070] In a specific example, Sentinel-1 synthetic aperture radar imagery data covering 19 eucalyptus plantation areas in Guangxi Zhuang Autonomous Region were acquired, with a time range from 2017 to 2024. The multi-temporal images were screened and organized according to a unified orbit type, imaging mode, and polarization method to construct a multi-year long-term series dataset with temporal continuity and spatial consistency.

[0071] S1.2 Uses digital elevation model to perform distance-Doppler terrain correction to eliminate geometric distortion caused by terrain undulation, ensure accurate registration of all temporal images in geospatial space, and avoid interference from terrain factors on subsequent change detection;

[0072] S1.3 employs a multi-temporal speckle filtering algorithm to reduce the inherent speckle noise in synthetic aperture radar (SAR) images and improve the data signal-to-noise ratio. In the specific example, the Refined Lee algorithm is selected as the speckle filtering method to reduce speckle noise in the radar image.

[0073] S1.4 performs radiometric calibration on the synthetic aperture radar image after speckle filtering, converting the radar digital quantization values ​​into backscattering coefficients. This enables the standardization and physical quantification of radiation values.

[0074] S2: Construct a time series of backscattering coefficients based on the preprocessed long-term time series dataset.

[0075] In one implementation, S2 includes:

[0076] S2.1: Arrange the images in the preprocessed multi-year long-term dataset in chronological order;

[0077] S2.2: For each pixel within the target area, extract its backscattering coefficient values ​​across all effective time phases to construct a backscattering coefficient time series.

[0078] Specifically, the constructed backscattering coefficient time series (pixel-level time-backscattering coefficient series) is represented as follows: ,in Representing the The backscattering coefficient of this pixel at each time phase. This represents the total number of time phases in the time series dataset, and the time span covered by this sequence is... (Start time) to (Termination time) records the complete changes in backscattering characteristics of a single pixel within the monitoring time range. In a specific example, to .

[0079] S3: Based on the rotation cycle and growth characteristics of short-cycle fast-growing plantations, a sliding time window parameter is set to divide the backscattering coefficient time series into sliding segments.

[0080] Specifically, the width of the sliding time window is determined as follows: Where L represents the length of time covered by a single sliding window, preferably 3 years.

[0081] Determine the window width parameter Based on this, set the step size parameter for the sliding time window. The moving step size This represents the translation interval between two adjacent sliding windows on the time axis, preferably 1 year.

[0082] Then, the time series is successively divided by sliding according to the window width and step size.

[0083] S4: Based on the partitioned sliding time window, construct the original difference sequence and the cumulative difference sequence.

[0084] In one implementation, S4 includes:

[0085] S4.1: Calculate the mean of all observations contained within the divided sliding time window, and use it as the baseline reference value for the un-logged state within the window;

[0086] S4.2: Calculate the deviation of each observation from the mean within the sliding time window and construct the original difference sequence;

[0087] S4.3: The original difference sequence is accumulated along the time sequence to obtain the cumulative difference sequence within the sliding time window. The curve shape of the cumulative difference sequence is used to reflect the structural change trend of the time series.

[0088] Specifically, S4.1: For any time window The window contains Each phase observation. Calculate all observations within the calculation window. mean This serves as a baseline reference value for the unfelled state within the window;

[0089] Step S4.2: Calculate each observation within the window. with the mean The degree of deviation is used to obtain the difference. This allows for the construction of the difference sequence corresponding to the window. ,in, The difference between the first observation and the mean. For the first The difference between each observation and the mean;

[0090] Step S4.3: For the difference sequence Perform term-by-term accumulation operations in the time domain to construct a cumulative sum sequence. The curve shape of this sequence reflects the structural change trend of the time series. If a logging event occurs within the window, the cumulative sum sequence will show a significant abrupt change in maximum value.

[0091] S5: Calculate the true range based on the constructed cumulative difference sum sequence, and extract the time point corresponding to the maximum value of the cumulative difference sum.

[0092] In one implementation, S5 includes:

[0093] S5.1: Based on the constructed cumulative sum of differences sequence, calculate the maximum and minimum values ​​in the cumulative sum sequence within the sliding time window, and calculate the true range based on the maximum and minimum values ​​in the cumulative sum sequence, which serves as the core feature for determining whether there is a structural change;

[0094] S5.2: The time point when the cumulative difference sum sequence reaches its maximum value is taken as the potential time point at which the logging event is most likely to occur within the sliding time window.

[0095] In the specific implementation process, S5.1 calculates the maximum value in the cumulative sum sequence within the window based on the cumulative sum sequence constructed in step S4. and minimum value The range was calculated. This serves as a core feature for determining whether structural changes exist;

[0096] S5.2: Determine if the cumulative sum of differences reaches its maximum value. The corresponding time point This point in time is the potential time point within the window where a logging event is most likely to occur.

[0097] S6: Randomly select a continuous subsequence of a preset length from the original difference sequence, and construct a resampled difference sequence using sampling with replacement.

[0098] In one implementation, S6 includes:

[0099] S6.1: A random block resampling strategy is adopted, and the length of the resampling block is set based on the temporal correlation of the backscattering coefficient time series.

[0100] S6.2: Using sampling with replacement, a continuous subsequence of a set length is randomly selected from the original difference sequence. The selected continuous subsequences are spliced ​​together in the order of selection until the length of the spliced ​​new sequence reaches or exceeds the window length.

[0101] S6.3: Extract a number of data points corresponding to the window length from the spliced ​​new sequence to generate a resampled difference sequence.

[0102] In the specific implementation process, S6.1 sets the length of the resampling block to be... The preferred duration is 30 days.

[0103] S6.2: Using sampling with replacement, from the original difference sequence The length of the random sample is A continuous subsequence (i.e., "random blocks"); the extracted random blocks are concatenated sequentially according to the extraction order until the length of the new concatenated sequence reaches or exceeds the window length. .

[0104] S6.3: Perform truncation on the concatenated sequence, extracting the unconcatenated portion. Data points, generating the first Resampled difference sequence .

[0105] Step 6, resampling based on random blocks, involves resampling the original difference sequence. Divided into several time-continuous sub-blocks (each sub-block has a length of...). When the number of observations in the window Cannot be During integer division, the last sub-block can be truncated or merged. Then, the original difference sequence is processed in sub-block units. According to a fixed length After segmentation, the resulting set of all candidate sub-blocks includes the standard sub-block (each in the original difference sequence). (composed sub-blocks) and end processing block (when the window length is...) Samples are randomly selected from the last special-length sub-block generated by truncation or merging. The selected sub-blocks are then sequentially concatenated until their length reaches or exceeds the window length. , cut off The data point generates the first... Resampled difference sequence This strategy maintains the overall randomness of the new sequence while inheriting the temporal characteristics of natural fluctuations under non-logging conditions. This makes the constructed range null hypothesis distribution more closely resemble the natural fluctuations under non-logging conditions in real-world noise environments, significantly reducing the false alarm rate under complex weather conditions and improving the robustness of the algorithm.

[0106] S7: Calculate the cumulative sum of resampled differences based on the resampled difference sequence, and construct the set of range null hypothesis distributions;

[0107] In one implementation, S7 includes:

[0108] S7.1: For the generated resampled difference sequence, perform an accumulation operation on the resampled difference sequence in chronological order to obtain the resampled difference cumulative sum sequence;

[0109] S7.2: The range is calculated based on the cumulative sum of resampled differences. The range obtained from multiple samplings and calculations constitutes the set of the null hypothesis distribution of the range.

[0110] In the specific implementation process, S7.1 addresses the resampled difference sequence generated in step S6. Recalculate according to the calculation logic of step S4 to construct the corresponding resampled difference cumulative sum sequence. ;

[0111] Step S7.2: Extract the first [item] according to the method in step S5. Simulated range .

[0112] S7.3: According to The resampling strategy repeats steps S6, S7.1, and S7.2. Next, construct by The set of range null hypothesis distributions constituted.

[0113] Specifically, the adaptive resampling strategy aims to balance test accuracy and computational efficiency. Given the large volume of remote sensing image data and the high computational load per pixel, it avoids redundant calculations within a fixed window. Specifically, it controls the number of resampling iterations through staged resampling and confidence level assessment. The dynamic determination is based on two pre-set sets of core parameters: the initial resampling count. Number of resampling operations with target (in, Simultaneously set a significance threshold. First execute Secondary resampling operation to calculate preliminary confidence level ,like If the window remains unchanged, the subsequent resampling calculation is terminated to save resources; if Then continue performing the resampling operation until the total number of times reaches [the target value]. Calculate the final confidence level In order to conduct accurate testing.

[0114] In specific implementation, considering that the Sentinel-1 satellite has a high-frequency revisit cycle of approximately 6 days, Number of valid observation samples within the sliding window of the year Typically around 180. This provides ample sample space for random block-guided resampling detection. Given the large data volume and high per-pixel computational load of remote sensing images, this invention designs... The principles for determining the value are as follows: Set the significance level. Based on the convergence analysis of the Monte Carlo simulation, in order to ensure... The relative error of the value estimation is controlled within an acceptable range, while also taking into account computational efficiency and the number of resampling operations. Set as (in For safety factors, a value of 10-20 is preferred. In this embodiment, it is preferred to set... This can provide of Value resolution, and can be compared with traditional This setting saves approximately 50% of computing resources, enabling rapid monitoring over a wide area.

[0115] S8: Based on the relationship between the true range and the set of distributions with the null hypothesis of the range, perform a statistical significance test on the candidate change points;

[0116] In one implementation, S8 includes:

[0117] S8.1: Compare the true range calculated based on the original data with the set of distributions of the null hypothesis of the range, and calculate the confidence level;

[0118] S8.2: If the confidence level is less than or equal to the set significance threshold, the judgment result is statistically significant and a significant change has occurred.

[0119] It is a statistic obtained from the null hypothesis distribution constructed based on random block resampling. Its physical meaning is: under the null hypothesis that "there is no structural change in the time series", the range obtained by random resampling is not less than the true range. The probability estimate is used to reflect the likelihood of an equal or greater magnitude of change than the actual observation, under conditions driven solely by random fluctuations or noise.

[0120] Significance threshold It is a pre-defined statistical criterion, its physical meaning being: the upper limit of the probability that a false rejection of the null hypothesis is allowed in a statistical test. It is used for interpretation and constraint. The criteria for judgment.

[0121] The relationship between the two can be summarized as follows: It is a statistical result quantity calculated based on data. This is a preset standard used to judge the result. If If the result is statistically significant, the null hypothesis of "no structural change" is rejected, meaning that a significant change has occurred.

[0122] In the specific implementation process, S8.1 will calculate the true range based on the original data. The confidence level is calculated by comparing the set of range null hypothesis distributions constructed in step S7 with the set of distributions constructed in step S7. .in, This is an indicator function that takes the value 1 when the condition is met, and 0 otherwise.

[0123] Step S8.2: Preset significance threshold (Preferred value: 0.05) Comparison confidence level Significance threshold If the calculated If the result is statistically significant, the null hypothesis of "no structural change" is rejected.

[0124] S9: Determine whether a candidate change point is a felling point based on the interval between the time points corresponding to the time points of the significance test results and the maximum value of the cumulative sum of differences.

[0125] In the specific implementation process, if the significance test result in step S8 meets the requirements, then the time point determined in step S5.2 is further determined. Is it in the current sliding window? Within the central year range, only when Only when the point falls within the central year range will it be marked as a definitive logging point. This strategy effectively avoids false detections caused by the sliding window edge effect.

[0126] Please see Figure 2 This is a schematic diagram of the logging detection results of a eucalyptus plantation in Guangxi Zhuang Autonomous Region. In the diagram, the horizontal axis represents time, the vertical axis represents the backscattering coefficient value of the corresponding pixel, the blue broken line is the backscattering time series of pixels in the target eucalyptus plantation area, and the red dashed line represents the change time points determined by the method of this embodiment.

[0127] Example 2

[0128] Based on the same inventive concept, this embodiment discloses a fast-growing forest deforestation detection device that takes into account the temporal autocorrelation of Sentinel-1. Please refer to [link to relevant documentation]. Figure 3 ,include:

[0129] The data acquisition and preprocessing module 201 is used to acquire multi-temporal Sentinel-1 synthetic aperture radar images of the target area, construct a multi-year long time series dataset with spatiotemporal consistency, and preprocess the constructed multi-year long time series dataset.

[0130] The backscattering coefficient time series construction module 202 is used to construct the backscattering coefficient time series based on the preprocessed multi-year long time series dataset;

[0131] The sliding time window division module 203 is used to set sliding time window parameters according to the rotation cycle and growth characteristics of short-cycle fast-growing plantations, and to divide the backscattering coefficient time series successively according to the set sliding time window parameters.

[0132] The difference cumulative sum sequence construction module 204 is used to construct the original difference sequence and the difference cumulative sum sequence based on the partitioned sliding time window;

[0133] The true range calculation and time point extraction module 205 is used to calculate the true range based on the constructed difference cumulative sum sequence, which serves as the core feature for determining whether there is a structural change, and to extract the time point corresponding to the maximum value of the difference cumulative sum.

[0134] The resampled difference sequence construction module 206 is used to randomly extract a continuous subsequence of a preset length from the original difference sequence and construct the resampled difference sequence using sampling with replacement.

[0135] The range null hypothesis distribution set construction module 207 is used to calculate the resampled difference cumulative sum sequence based on the resampled difference sequence and construct the range null hypothesis distribution set;

[0136] The significance test module 208 is used to perform statistical significance tests on candidate change points based on the relationship between the true range and the set of range null hypothesis distributions.

[0137] The logging time point location module 209 is used to determine whether a candidate change point is a logging point based on the interval of the time point corresponding to the time point of the significance test result and the maximum value of the cumulative sum of differences.

[0138] Since the device in Embodiment 2 of this invention is the same device used in the fast-growing forest felling detection method considering the time-series autocorrelation of Sentinel-1 in Embodiment 1, those skilled in the art can understand the specific structure and variations of this device based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All devices used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0139] Example 3

[0140] Based on the same inventive concept, the present invention also provides a computer device, including 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 method described in Embodiment 1.

[0141] Since the computer device described in Embodiment 3 of this invention is the same computer device used in implementing the fast-growing forest felling detection method considering Sentinel-1 time-series autocorrelation in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer device based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All computer devices used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0142] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0143] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0144] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various modifications and variations to the embodiments of the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the embodiments of the invention fall within the scope of the claims of the invention and their equivalents, the invention also intends to include these modifications and variations.

Claims

1. A method for detecting logging in fast-growing forests that takes into account the temporal autocorrelation of Sentinel-1, characterized in that, include: Acquire multi-temporal Sentinel-1 synthetic aperture radar images of the target area, construct a multi-year long-time series dataset with spatiotemporal consistency, and preprocess the constructed multi-year long-time series dataset; Based on the preprocessed multi-year long-time series dataset, a time series of backscattering coefficients is constructed. Based on the rotation cycle and growth characteristics of short-cycle fast-growing plantations, a sliding time window parameter was set to divide the backscattering coefficient time series into sliding segments. Based on the partitioned sliding time window, construct the original difference sequence and the cumulative difference sequence; The true range is calculated based on the constructed cumulative difference sum sequence, and the time point corresponding to the maximum value of the cumulative difference sum is extracted. A continuous subsequence of a preset length is randomly selected from the original difference sequence, and a resampled difference sequence is constructed using sampling with replacement. Calculate the cumulative sum of resampled differences based on the resampled difference sequence, and construct the set of range null hypothesis distributions; Based on the relationship between the true range and the set of distributions with the null hypothesis of the range, a statistical significance test is performed on the candidate change points; Based on the significance test results and the time interval corresponding to the maximum value of the cumulative sum of differences, determine whether the candidate change point is a felling point.

2. The fast-growing forest deforestation detection method considering the temporal autocorrelation of Sentinel-1 as described in claim 1, characterized in that, Preprocessing of the constructed multi-year long-term time series dataset includes: Standardize the image data in the multi-year long-time series dataset; Distance-Doppler terrain correction using digital elevation models; A multi-temporal speckle filtering algorithm is used to reduce the inherent speckle noise in synthetic aperture radar images; Convert radar digital quantization values ​​into backscattering coefficients.

3. The fast-growing forest deforestation detection method considering the temporal autocorrelation of Sentinel-1 as described in claim 1, characterized in that, Based on the preprocessed multi-year long-term time series dataset, a time series of backscattering coefficients is constructed, including: Arrange the images in the preprocessed multi-year long-time series dataset in chronological order; For each pixel within the target area, its backscattering coefficient values ​​are extracted across all effective time phases to construct a backscattering coefficient time series.

4. The fast-growing forest logging detection method considering the temporal autocorrelation of Sentinel-1 as described in claim 1, characterized in that, Based on the partitioned sliding time window, the original difference sequence and the cumulative difference sequence are constructed, including: The mean of all observations included in the divided sliding time window is calculated and used as the baseline reference value for the un-logged state within the window; Calculate the deviation of each observation from the mean within the sliding time window, and construct the original difference sequence; The original difference sequence is accumulated along the time sequence to obtain the cumulative difference sequence within the sliding time window. The curve shape of the cumulative difference sequence is used to reflect the structural change trend of the time series.

5. The fast-growing forest logging detection method considering the temporal autocorrelation of Sentinel-1 as described in claim 1, characterized in that, The true range is calculated based on the constructed cumulative sum of differences sequence, and the time point corresponding to the maximum value of the cumulative sum of differences is extracted, including: Based on the constructed cumulative sum of differences sequence, the maximum and minimum values ​​in the cumulative sum sequence within the sliding time window are calculated. The true range is then calculated based on the maximum and minimum values ​​in the cumulative sum sequence, serving as the core feature for determining whether structural changes exist. The time point at which the cumulative difference sequence reaches its maximum value is taken as the potential time point at which the logging event is most likely to occur within the sliding time window.

6. The fast-growing forest deforestation detection method considering the temporal autocorrelation of Sentinel-1 as described in claim 1, characterized in that, A pre-defined continuous subsequence is randomly selected from the original difference sequence, and a resampled difference sequence is constructed using sampling with replacement, including: A random block resampling strategy is adopted, and the length of the resampling block is set based on the temporal correlation of the backscattering coefficient time series. Using sampling with replacement, a continuous subsequence of a set length is randomly selected from the original difference sequence. The selected continuous subsequences are then spliced ​​together in the order of selection until the length of the spliced ​​new sequence reaches or exceeds the window length. Extract the number of data points corresponding to the window length from the spliced ​​new sequence to generate a resampled difference sequence.

7. The fast-growing forest logging detection method considering the temporal autocorrelation of Sentinel-1 as described in claim 1, characterized in that, Calculate the cumulative sum of resampled differences based on the resampled difference sequence, and construct a set of range null hypothesis distributions, including: For the generated resampled difference sequence, the resampled difference sequence is accumulated along the time sequence to obtain the resampled difference cumulative sum sequence; The range is calculated based on the cumulative sum of resampled differences. The range obtained from multiple samplings and calculations constitutes the set of the null hypothesis distribution of the range.

8. The fast-growing forest logging detection method considering the temporal autocorrelation of Sentinel-1 as described in claim 1, characterized in that, Based on the relationship between the true range and the set of distributions with the null hypothesis of the range, statistical significance tests are performed on the candidate change points, including: The confidence level is calculated by comparing the true range calculated based on the original data with the set of distributions of the null hypothesis of the range. If the confidence level is less than or equal to the set significance threshold, the result is considered statistically significant, indicating a significant change.

9. A fast-growing forest deforestation detection device that takes into account the temporal autocorrelation of Sentinel-1, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-temporal Sentinel-1 synthetic aperture radar images of the target area, construct a multi-year long-time series dataset with spatiotemporal consistency, and preprocess the constructed multi-year long-time series dataset. The backscattering coefficient time series construction module is used to construct backscattering coefficient time series based on preprocessed multi-year long time series datasets. The sliding time window division module is used to set sliding time window parameters based on the rotation cycle and growth characteristics of short-cycle fast-growing plantations, and to divide the backscattering coefficient time series successively according to the set sliding time window parameters. The difference cumulative sum sequence construction module is used to construct the original difference sequence and the difference cumulative sum sequence based on the partitioned sliding time window; The module for calculating the true range and extracting time points is used to calculate the true range based on the constructed cumulative difference sum sequence, which serves as the core feature for determining whether there is a structural change, and to extract the time point corresponding to the maximum value of the cumulative difference sum. The resampled difference sequence construction module is used to randomly extract a continuous subsequence of a preset length from the original difference sequence and construct the resampled difference sequence using sampling with replacement. The range null hypothesis distribution set construction module is used to calculate the cumulative sum of resampled differences based on the resampled difference sequence and construct the range null hypothesis distribution set; The significance test module is used to perform statistical significance tests on candidate change points based on the relationship between the true range and the set of distributions of the null hypothesis of the range. The logging time point location module is used to determine whether a candidate change point is a logging point based on the interval of the time point corresponding to the time point of the cumulative sum of differences and the significance test results.

10. A computer device, 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 fast-growing forest logging detection method that takes into account the timing autocorrelation of Sentinel-1 as described in any one of claims 1 to 8.

Images