A method for determining snow disaster-affected areas in bamboo forests based on a GEE platform

Abstract

This invention relates to the field of remote sensing image recognition technology, and in particular to a method for determining snow disaster-affected areas of bamboo forests based on a GEE platform. This method primarily addresses the problems of low efficiency, high cost, and difficulty in achieving large-scale rapid assessment caused by reliance on manual field surveys in existing technologies. This invention is achieved through the following technical solution: First, multi-temporal Sentinel-2 images before and after the snow disaster are acquired and preprocessed; then, spectral bands, vegetation indices, and texture features are extracted to form an initial feature set; next, key feature variable combinations are determined through statistical significance filtering and machine learning optimization; based on this, a logistic regression discriminant model is established and the optimal classification threshold is determined; finally, the trained model is used to classify the snow disaster status of the bamboo forest area, generating a spatial distribution map of the affected area.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image recognition technology, and in particular to a method for determining snow disaster-affected areas of bamboo forests based on a GEE platform. Background Technology

[0002] Snow and ice disasters are a significant abiotic factor affecting the stability of moso bamboo forest ecosystems, directly impacting bamboo growth. Mild effects include significant bending of bamboo stalks, while severe effects include breakage, uprooting, and death – losses that are irreversible. Previous research has shown that among many subtropical forest tree species, moso bamboo is the most severely affected by snow and ice disasters.

[0003] Currently, the main method for determining the affected areas of bamboo forests after snow disasters relies on traditional ground-based field surveys. This method involves manually setting up sample plots deep within the forest area and counting the number of affected bamboo plants to calculate the damage rate. This method has significant drawbacks: First, it is extremely costly in terms of manpower, resources, and time, resulting in very low survey efficiency. Second, bamboo forests are mostly distributed in mountainous areas with complex terrain and poor post-disaster road conditions, making many areas inaccessible and creating numerous blind spots in the survey results, failing to fully reflect the true spatial distribution of damage. Third, this method is highly dependent on the subjective experience of the survey personnel; different individuals may have different standards for judging the degree of damage, leading to poor comparability and accuracy of the survey results.

[0004] While satellite remote sensing technology offers advantages such as wide-area, periodic observation, providing new avenues for disaster monitoring (e.g., using spectral indices like the Normalized Difference Vegetation Index to reflect vegetation growth), existing remote sensing methods still have several shortcomings when practically applied to monitoring snow disasters in bamboo forests. On one hand, single vegetation indices are insensitive to structural changes in bamboo forests caused by snow disasters, such as patchy damage and lodging, making it difficult to accurately capture disaster characteristics. On the other hand, although numerous remote sensing features are available, the lack of effective feature selection methods leads to high feature redundancy and significant noise interference, making it impossible to pinpoint the core indicators most closely related to snow disasters in bamboo forests. Furthermore, existing models often directly apply general algorithms, failing to construct dedicated discriminative models and optimal thresholds for the unique response mechanisms of snow disasters in bamboo forests, resulting in limited classification accuracy and weak practicality. Summary of the Invention

[0005] The purpose of this invention is to provide a method for determining the snow disaster-affected area of ​​bamboo forests based on the GEE platform, so as to solve the problems in the background art.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for determining snow disaster-affected areas in bamboo forests based on a GEE platform includes the following steps:

[0008] S1: Using the GEE platform and in conjunction with preset cloud cover thresholds and time windows, radiometric calibration and atmospheric correction are performed on Sentinel-2 image sequences at specific time phases before and after the snow disaster to generate a pre-processed set of multi-temporal remote sensing images with physical significance of surface reflectance before and after the disaster.

[0009] S2: Using the pre-disaster and post-disaster multi-temporal remote sensing image set, multi-dimensional feature extraction covering spectral band reflectance, multiple vegetation indices and NDVI-derived texture features is completed for the two temporal phases before and after the disaster, generating an initial feature set containing all features of the pre-disaster and post-disaster states.

[0010] S3: For the initial set of features, by calculating the difference between the post-disaster feature values ​​and the pre-disaster feature values ​​and implementing significance filtering based on statistical hypothesis testing, a subset of disaster feature variables that only include those that have changed significantly before and after the snow disaster is obtained.

[0011] S4: Based on the subset of catastrophic feature variables, redundancy elimination based on correlation analysis and feature importance screening based on multiple machine learning algorithms are carried out in sequence, and a combination of key feature variables consisting of core spectral bands, vegetation index and texture features is output.

[0012] S5: Using the disaster information represented by the combination of key feature variables, perform parameter training of the logistic regression model and evaluate the model's discrimination performance with ROC curves to establish a snow disaster discrimination model with a clear mathematical expression and its optimal classification threshold that can best distinguish between disaster-affected and unaffected pixels.

[0013] S6: Apply the snow disaster discrimination model and its optimal classification threshold to classify the snow disaster status of each pixel in the distribution area of ​​all bamboo forests in the target area, and output a spatial distribution map of the snow disaster-affected area of ​​bamboo forests in the form of a binary image.

[0014] Preferably, in step S1:

[0015] The preset time window is the cloudless image period during the peak vegetation growth season within one year before the snow disaster, and the image period after the snow disaster when the influence of clouds and snow has basically subsided.

[0016] The preset cloud cover threshold is set to a predetermined percentage threshold for the entire image;

[0017] Radiometric calibration and atmospheric correction were performed using the Sen2Cor algorithm embedded in the GEE platform;

[0018] The generated pre-disaster and post-disaster multi-temporal remote sensing image sets have consistent spatial resolution and geographic coordinate reference.

[0019] Preferably, in step S2:

[0020] The spectral reflectance includes at least the B4 red band, B7 red edge 3 band, B8 near-infrared band and B11 short-wave infrared 1 band of Sentinel-2.

[0021] The vegetation indices include at least the atmospheric resistance vegetation index (ARVI), normalized water index (NDMI), red-edged normalized vegetation index (NDRE1), normalized water index (NDWI), and ratio vegetation index (RVI).

[0022] NDVI-derived texture features are homogeneous texture features calculated from NDVI images using a gray-level co-occurrence matrix.

[0023] The initial feature set contains all feature variables calculated in both the pre-disaster and post-disaster time phases.

[0024] Preferably, in step S3:

[0025] The calculation of the difference between the post-disaster characteristic value and the pre-disaster characteristic value specifically involves subtracting the pre-disaster characteristic value from the post-disaster characteristic value to obtain the disaster difference value for each characteristic.

[0026] Significance filtering based on statistical hypothesis testing, specifically using the two-sample t-test method;

[0027] The t-test is set with a predefined significance level to screen out characteristic variables that change significantly;

[0028] The obtained subset of catastrophic feature variables forms the basis for subsequent feature selection processes.

[0029] Preferably, in step S4, the redundancy elimination based on correlation analysis specifically includes:

[0030] Calculate the Pearson correlation coefficients between all pairs of variables in the subset of catastrophic characteristic variables;

[0031] Set a predefined threshold for determining high correlation;

[0032] For variable pairs whose absolute correlation coefficient exceeds this threshold, calculate the average absolute correlation coefficient between each variable and all other variables in the subset;

[0033] The variable with the highest absolute value of the average correlation coefficient is removed to retain features that are more independent and less redundant.

[0034] Preferably, in step S4, the feature importance screening based on multiple machine learning algorithms specifically includes:

[0035] LASSO regression analysis was used, and L1 regularization was used to force the regression coefficients to shrink, compressing the coefficients of unimportant variables to zero, thereby achieving feature screening.

[0036] At the same time, the random forest algorithm is used to rank the features by calculating the average reduction in impurity of the features during the construction of multiple decision trees;

[0037] The intersection of the variables selected by both LASSO regression analysis and random forest algorithm is taken as the final combination of key feature variables.

[0038] The key feature variable combination consists of eight variables: homogeneous texture features of NDVI, B4 red band, B7 red edge 3 band, ARVI, NDMI, NDRE1, NDWI, and RVI.

[0039] Preferably, the variables in the combination of key feature variables have clear physical meanings and calculation formulas:

[0040] The formula for calculating the Atmospheric Resistant Vegetation Index (ARVI) is as follows: ;

[0041] The formula for calculating the Normalized Diurnal Moisture Index (NDMI) is: ;

[0042] The formula for calculating the Red Edge Normalized Difference Vegetation Index (NDRE1) is as follows: ;

[0043] The formula for calculating the Ratio Vegetation Index (RVI) is: .

[0044] Preferably, in step S5:

[0045] The logistic regression model is trained by the maximum likelihood estimation method, and the discriminant equation it establishes is a weighted sum of a linear function containing the combination of the key feature variables and a constant term.

[0046] The ROC curve is used to evaluate the performance of the discriminant equation at different classification thresholds;

[0047] The optimal classification threshold is determined based on the analysis results of the ROC curve to balance the classification accuracy, sensitivity and specificity of the model.

[0048] Preferably, in step S6:

[0049] The snow disaster status of each pixel in the distribution area of ​​all bamboo forests in the target area is classified one by one. The specific criterion is: calculate the predicted value y of each pixel. If the y value is greater than the optimal classification threshold, the pixel is determined to be in a disaster state, otherwise it is in an undisaster state.

[0050] In a binary image, different pixel values ​​are used to represent disaster-stricken and unaffected areas, respectively.

[0051] The output spatial distribution map of the snow disaster-affected area of ​​the bamboo forest contains geographic coordinate information, which can be directly loaded into GIS software for area measurement and spatial analysis.

[0052] The beneficial effects of this invention are:

[0053] This invention fully utilizes the multi-source remote sensing data and powerful computing capabilities of the GEE cloud platform to achieve fully automated processing of large areas, completely changing the traditional work mode that relies on manual on-site surveys. This not only avoids the limitations imposed by complex terrain on survey work, but also greatly saves manpower, material resources, and time costs, making rapid and frequent disaster monitoring and assessment possible.

[0054] Unlike traditional methods that rely on a single vegetation index or empirical feature selection, this invention constructs and implements a rigorous multi-step feature screening process. This process can accurately identify the core indicator combination most relevant to snow disaster response from a massive amount of initial features. This technical solution effectively overcomes feature redundancy and noise interference, enabling the final discriminative model to more sensitively capture the unique physiological and structural changes in bamboo forests after a snow disaster, thereby achieving accurate differentiation of affected areas.

[0055] This invention does not simply apply a general algorithm, but rather constructs a dedicated logistic regression discriminant model and determines the optimal classification threshold, forming a stable and reliable set of classification decision rules. Its final output is an intuitive spatial distribution map of the disaster-stricken area, providing forestry management departments with clear decision-making support. It strongly supports post-disaster relief deployment, loss assessment, and the sustainable management of bamboo forests, demonstrating high practical value and promising prospects for wider application. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a method for determining snow disaster-affected areas of bamboo forests based on a GEE platform according to the present invention. Detailed Implementation

[0057] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0058] Example: Refer to Figure 1 As shown in this embodiment, a method for determining the snow disaster-affected area of ​​bamboo forest based on a GEE platform includes the following steps:

[0059] S1: Using the GEE platform and in conjunction with preset cloud cover thresholds and time windows, radiometric calibration and atmospheric correction are performed on Sentinel-2 image sequences at specific time phases before and after the snow disaster to generate a pre-processed set of multi-temporal remote sensing images with physical significance of surface reflectance before and after the disaster.

[0060] S2: Using the pre-disaster and post-disaster multi-temporal remote sensing image set, multi-dimensional feature extraction covering spectral band reflectance, multiple vegetation indices and NDVI-derived texture features is completed for the two temporal phases before and after the disaster, generating an initial feature set containing all features of the pre-disaster and post-disaster states.

[0061] S3: For the initial set of features, by calculating the difference between the post-disaster feature values ​​and the pre-disaster feature values ​​and implementing significance filtering based on statistical hypothesis testing, a subset of disaster feature variables that only include those that have changed significantly before and after the snow disaster is obtained.

[0062] S4: Around the subset of catastrophic feature variables, redundancy elimination based on correlation analysis and feature importance optimization based on multiple machine learning algorithms are carried out in sequence, and a key feature variable combination consisting of core spectral bands, vegetation index and texture features is output.

[0063] S5: Using the disaster information represented by the combination of key feature variables, perform parameter training of the logistic regression model and evaluate the model's discrimination performance with ROC curves to establish a snow disaster discrimination model with a clear mathematical expression and its optimal classification threshold that can best distinguish between disaster-affected and unaffected pixels.

[0064] S6: Apply the snow disaster discrimination model and its optimal classification threshold to classify the snow disaster status of each pixel in the distribution area of ​​all bamboo forests in the target area, and output a spatial distribution map of the snow disaster-affected area of ​​bamboo forests in the form of a binary image.

[0065] In this embodiment of the invention, the GEE platform is used in conjunction with a preset cloud cover threshold and time window to perform radiometric calibration and atmospheric correction on Sentinel-2 image sequences at specific time points before and after a snow disaster, generating a preprocessed multi-temporal remote sensing image set with physical significance of surface reflectance before and after the disaster. Specific implementation methods include:

[0066] This invention uses a concentrated bamboo forest distribution area in Shengzhou City, Zhejiang Province as the target monitoring area. In the GEE platform, the study area is defined by uploading a vector boundary file of the region or directly using latitude and longitude coordinates. The data source is Sentinel-2 Multispectral Imager Level-1C data provided by the European Space Agency. The core reasons for choosing this data source are its high spatial resolution, short revisit period, free and open access to the data, and its multispectral characteristics, covering visible light, red-edge, near-infrared, and short-wave infrared, which can effectively capture the physiological and structural information of vegetation, making it very suitable for fine-grained monitoring of vegetation dynamics at a regional scale.

[0067] For specific snow disaster events, accurately defining the "pre-disaster" and "post-disaster" time phases is a prerequisite for accurately capturing the disaster's impact. In this embodiment, the pre-disaster time phase selects images from the peak vegetation growth season within the year prior to the snow disaster to ensure that the bamboo forest is in its most physiologically active state, and its spectral characteristics can represent a healthy, undisturbed background level. The post-disaster time phase selects images from after the snow disaster, when the surface snow has completely melted and the weather is mainly clear, to minimize the interference of snow cover and residual cloud cover on the true surface reflection information. To achieve automated batch filtering, the following method is used in the GEE code editor: Interface, through The method takes a preset pre-disaster and post-disaster time window as input and utilizes... Define the research area. Most importantly, through... By setting a cloud cover threshold, the platform will automatically filter out invalid images with excessively high cloud cover, ensuring the availability of the constructed dataset.

[0068] Sentinel-2L1C data provides geometrically corrected top-surface atmospheric reflectance, which still includes atmospheric scattering and absorption effects. Atmospheric correction is necessary to obtain the true surface reflectance. This embodiment of the invention directly calls and utilizes the embedded [feature / technology] within the GEE platform. The dataset consists of L2A-level surface reflectance products generated by the ESA's sen2cor algorithm. This process is physically achieved by solving the atmospheric radiative transfer equation, the core of which is converting the apparent reflectance received by the sensors into the true surface reflectance. This conversion process can be expressed by the following simplified physical model: ,in: It is the bidirectional reflectance of the target surface at wavelength λ, which is the ultimate target value pursued in this step; It is the apparent reflectance received at the satellite sensor, i.e., the input of L1C data; It is the atmospheric path radiation term, which represents the radiation component that directly enters the sensor's field of view after being scattered by atmospheric molecules and aerosols. and These are the transmittance caused by specific gas absorption and atmospheric molecule scattering, respectively, and the transmittance relative to the solar zenith angle ( ). ) and observed zenith angle ( Related to; The contribution is from environmental irradiance, mainly from the cross-radiation effect generated after neighboring pixels are scattered by the atmosphere.

[0069] The sen2cor algorithm automatically performs the aforementioned complex physical calculations by integrating digital elevation model data and utilizing imagery information to estimate atmospheric composition. In this embodiment, developers only need to call preprocessed L2A data to efficiently and in batches obtain images with physical significance of surface reflectance, laying a solid foundation for subsequent quantitative remote sensing analysis.

[0070] After undergoing the rigorous time-space filtering and physical correction described above, multiple valid pre-disaster and post-disaster images are composited into high-quality composite images representing their respective time phases, with no or very few clouds. In GEE, for all image sets within the same time phase, a... The method calculates the median for each pixel across all valid images. This median composite method effectively suppresses residual thin clouds, cloud shadows, and noise, and compensates for potential local data gaps in single images, ultimately generating two high-quality composite images representing the "pre-disaster background state" and the "post-disaster damage state." This yields two core data products: a "pre-disaster multi-temporal remote sensing image set" and a "post-disaster multi-temporal remote sensing image set." These two datasets share identical geographic coordinate references, spatial resolutions, and surface reflectance dimensions, ensuring direct comparability in subsequent change detection and feature calculation. These preprocessed image sets can be directly stored in the user's GEE assets for seamless use by subsequent feature extraction modules.

[0071] In this embodiment of the invention, the pre-disaster and post-disaster multi-temporal remote sensing image set is used to perform multi-dimensional feature extraction covering spectral band reflectance, multiple vegetation indices, and NDVI-derived texture features for both pre-disaster and post-disaster temporal phases, generating an initial feature set containing all features of both pre-disaster and post-disaster states. Specific implementation methods include:

[0072] Spectral reflectance is the most direct feature for remote sensing identification of ground features. Different bands exhibit specific response characteristics to chlorophyll content, water status, and cell structure of vegetation. This embodiment of the invention selects nine core bands closely related to vegetation monitoring from Sentinel-2 imagery for calculation. These bands include: B2: blue light, 490nm; B3: green light, 560nm; B4: red light, 665nm; B5: red edge 1, 705nm; B6: red edge 2, 740nm; B7: red edge 3, 783nm; B8: near-infrared, 842nm; B11: shortwave infrared 1, 1610nm; and B12: shortwave infrared 2, 2190nm. In the GEE platform, this is achieved by calling... The method directly extracts the surface reflectance values ​​of each pixel in the aforementioned bands from composite images of pre- and post-disaster data. The red band B4 is the main absorption region for chlorophyll, the near-infrared band B8 is sensitive to vegetation cell structure, while the short-wave infrared bands B11 and B12 are highly correlated with vegetation moisture content. Extracting reflectance from two bands at the same time points lays the foundation for subsequent calculations of spectral changes caused by the disaster.

[0073] Vegetation indices, by combining reflectance from different spectral bands, can amplify vegetation information and suppress background noise, thus more sensitively characterizing vegetation growth and biophysical parameters. This invention, based on the aforementioned spectral bands, systematically calculates twelve widely used vegetation indices, forming an index library reflecting different aspects of vegetation. These indices include, but are not limited to: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Normalized Difference Moisture Index (NDMI), Atmospheric Impedance Vegetation Index (ARVI), and Normalized Red Edge Index (NDRE). The calculation of each index is achieved through pixel-level algebraic operations in GEE. For example, the calculation formulas for the core NDVI and ARVI are as follows: , Among them, NDVI is a classic indicator reflecting vegetation greenness and photosynthetic intensity, while ARVI, by introducing the blue light band B2 to perform atmospheric correction on the red light band B4, reduces the impact of aerosol scattering and is more stable in situations where post-disaster atmospheric conditions may be unfavorable. All indices use... The function is computed in parallel in GEE to generate vegetation index layers for both pre-disaster and post-disaster time phases.

[0074] Snow disasters not only affect the physiological activity of bamboo forests but also alter their spatial distribution patterns, leading to "patchy" phenomena caused by lodging and breakage. Simple spectral information is insensitive to such textural changes. Therefore, this invention introduces texture analysis based on the gray-level co-occurrence matrix. In practice, the calculated NDVI imagery is first used as input data. In GEE, using... The method involves traversing the entire NDVI image using a sliding window, calculating the GLCM within each window, and extracting eight key texture features, including homogeneity, contrast, entropy, and energy. Taking the core homogeneity as an example, its calculation formula is as follows: ,in, These are the elements of the normalized gray-level co-occurrence matrix, where i and j are gray levels. A higher homogeneity value indicates a more uniform image texture. If sporadic lodging occurs in a post-disaster bamboo forest, it will create a heterogeneous pattern in the NDVI image, leading to reduced homogeneity and increased contrast and entropy. This step expands single spectral information to two-dimensional spatial contextual information, greatly enhancing the ability to identify damage to forest stand structure.

[0075] After completing the above three calculations, a total of 29 feature variables (9 bands + 12 vegetation indices + 8 texture features) were generated for each pixel in both the pre-disaster and post-disaster time phases. In GEE, using... The function stacks all feature layers into a multi-band image, forming "pre-disaster feature image" and "post-disaster feature image" respectively. At this point, the initial feature set is complete. This set comprehensively captures multi-dimensional and multi-scale state information of bamboo forests before and after the snow disaster, from micro-physiological aspects (spectral and indices) to macro-structural aspects (texture), providing sufficient data support for the next step of accurately locating snow disaster-sensitive indicators through differential analysis and statistical screening. The entire process is automated in GEE, ensuring efficiency and consistency when processing large areas.

[0076] In this embodiment of the invention, for the initial set of features, by calculating the difference between post-disaster feature values ​​and pre-disaster feature values ​​and implementing significance filtering based on statistical hypothesis testing, a subset of disaster feature variables containing only those that have changed significantly before and after the snow disaster is obtained. Specific implementation methods include:

[0077] In this embodiment of the invention, selecting disaster-related feature variables sensitive to snow disasters from the initial set of multi-dimensional features is a key preprocessing step in constructing a high-precision discriminative model. The goal of this stage is to eliminate redundant information and retain feature variables that are significantly related to the snow disaster response, providing high-quality input for subsequent modeling. The specific implementation includes two core steps: feature difference calculation and statistical significance filtering.

[0078] The initial feature set contains a complete set of features from both the pre- and post-disaster phases, but directly using absolute values ​​introduces background noise. To accurately capture the true changes caused by the snow disaster, this invention performs a difference operation between the post-disaster and pre-disaster values ​​for each feature variable. In the GEE platform, this process is efficiently implemented through pixel-level algebraic operations. Specifically, assuming... Images representing pre-disaster characteristics Images representing post-disaster characteristics, such as disaster difference images. The calculation formula is: .

[0079] This operation generates a disaster difference feature set with the same dimensions as the initial feature set, where the value of each pixel represents the specific change in that feature before and after the snow disaster. Positive values ​​indicate enhanced features after the disaster, while negative values ​​indicate weakened features. For example, a negative difference in NDVI may indicate a decrease in vegetation chlorophyll content or a decline in canopy cover; while a positive difference in shortwave infrared reflectance may reflect a deterioration in vegetation moisture conditions. This differential processing effectively eliminates the influence of common background factors such as seasonal variations and phenological cycles, purifying and enhancing the disaster signal.

[0080] The calculated catastrophic difference feature set may contain a large number of small fluctuations caused by random noise, sensor errors, or other non-catastrophic factors. To screen out the feature variables that are truly significant in response to snow disasters, this invention introduces a two-sample t-test for statistical significance assessment. In specific implementation, a sufficient number of validation sample points are first evenly distributed within the study area. In this embodiment, 200 sample points were distributed, half of which are located in the disaster-affected areas confirmed by on-site verification, and the other half are located in the unaffected areas. By extracting these sample points... For each feature variable, we constructed two sample sets: a difference sample from the affected group and a difference sample from the unaffected group. Then, we performed the following t-test on each feature variable: ,in and These are the sample means of the differences between the affected group and the unaffected group, respectively. and For sample size, To pool the standard deviations, the null hypothesis was that "there is no significant difference in the mean difference between the affected and unaffected groups for this feature." We set the significance level α to 0.05. If the calculated p-value is less than this threshold, we reject the null hypothesis, considering the change in this feature before and after the snow disaster to be statistically significant and therefore retain it.

[0081] The above process is fully automated within the GEE platform through JavaScript algorithm scripts. The algorithm traverses every feature band in the disaster difference feature set, automatically performing sample value extraction, t-statistic calculation, and p-value determination. All feature variables that pass the significance test are retained and integrated into a new, reduced-dimensional subset of disaster feature variables. Each feature in this subset has undergone rigorous statistical validation, confirming that its changes are significantly correlated with snow disaster events, rather than being random fluctuations. This step significantly compresses the feature space, effectively improving the training efficiency and generalization ability of subsequent machine learning models, laying a solid foundation for building a robust snow disaster discrimination model. The entire process embodies the feature selection philosophy of shifting from "data-driven" to "evidence-driven," ensuring the scientific rigor and reliability of the constructed disaster indicator system.

[0082] In this embodiment of the invention, redundancy elimination based on correlation analysis and feature importance screening based on multiple machine learning algorithms are sequentially performed around the subset of catastrophic feature variables, outputting a key feature variable combination composed of core spectral bands, vegetation index, and texture features. Specific implementation methods include:

[0083] Even after passing the significance test, highly correlated features may still exist, leading to information redundancy and increasing the risk of model overfitting. To eliminate redundancy, this invention calculates the Pearson correlation coefficient between all pairs of variables in the catastrophic feature variable subset. This coefficient measures the strength and direction of the linear correlation between variables, and its calculation formula is as follows: ,in, and This represents the value of two different feature variables at the i-th sample point. and These are the sample means for each variable, and n is the sample size. In the GEE platform, a script is used to automatically calculate the correlation coefficient matrix for all variable pairs. A threshold is set. Variable pairs exceeding this threshold are marked as highly correlated. For each highly correlated variable pair, the absolute value of its average correlation coefficient with all other variables in the subset is calculated, and the variable with the higher average correlation coefficient is removed. This step preserves more independent and representative features, forming a redundant intermediate feature subset, laying a cleaner data foundation for subsequent machine learning optimization.

[0084] To further filter the most discriminative features from the deduplicated intermediate feature subset, this invention employs two algorithms with different mechanisms—LASSO regression and random forest—for cross-validation. For LASSO regression, an L1 regularization term is added to the loss function of standard linear regression, forcing the coefficients of unimportant features to shrink to zero, thereby achieving feature selection. Its objective function is: ,in These are sample labels. It is an eigenvalue. It is a coefficient. It is a hyperparameter that controls the strength of regularization. The optimal one is selected through cross-validation. Features with non-zero coefficients are retained. Simultaneously, a random forest algorithm is run in parallel. This algorithm constructs multiple decision trees through bootstrap sampling and calculates the average reduction in impurity for each feature across all trees due to split nodes; this is known as Gini importance or average reduction in impurity, and is used as a measure of feature importance. Finally, the intersection of the variables selected by both LASSO regression and random forest methods is chosen as the final combination of key feature variables.

[0085] The aforementioned correlation analysis and machine learning optimization processes are all automatically completed within the GEE platform through integrated algorithm scripts. The script first performs correlation analysis and outputs a deredundant subset. This subset is then simultaneously input into both LASSO and Random Forest modules for feature importance evaluation and ranking. Finally, the intersection of the two results is taken. This rigorous three-step selection process ensures the final output combination of key feature variables, which in this embodiment is determined as... The eight variables are not only significantly correlated with snow disaster response, but also exhibit low redundancy among themselves and have been consistently recognized by various machine learning algorithms. This combination provides the most refined and effective input feature set for subsequently building a high-performance, highly interpretable snow disaster discrimination model.

[0086] In this embodiment of the invention, the disaster information represented by the combination of key feature variables is used to perform parameter training of a logistic regression model and evaluate the model's discriminative performance using ROC curves. This establishes a snow disaster discrimination model with a clear mathematical expression and its optimal classification threshold for distinguishing between affected and unaffected pixels. Specific implementation methods include:

[0087] In this embodiment of the invention, constructing a snow disaster discrimination model based on a combination of key feature variables and determining the optimal classification threshold are the core steps in transforming feature information into reliable classification decisions. This process establishes a complete classification system with a clear mathematical expression and optimal discrimination criteria through two stages: logistic regression model training and ROC curve analysis.

[0088] Logistic regression models are suitable for binary classification problems, capable of linearly combining input features and mapping them to disaster probability via the Sigmoid function, providing an ideal modeling framework for this study. In the GEE platform, a pre-prepared training sample set containing labeled "disaster-affected" and "non-disaster-affected" samples is used, along with eight finalized key feature variables. The catastrophic differences of B4, B7, ARVI, NDMI, NDRE1, NDWI, and RVI are correlated. The model solves for the optimal parameters using the maximum likelihood estimation method, the basic form of which is: first, a linear combination is constructed. ,in For the intercept term, arrive These are the regression coefficients for the eight characteristic variables. arrive These are the corresponding eigenvalues. Then, the Sigmoid function is applied. Convert linear combinations into probability values. In GEE, use... The classifier is trained using an iterative reweighted least squares method to optimize the parameters, ultimately yielding the specific discriminant equation: This equation fully characterizes the quantitative contribution of each feature variable to the snow disaster status of bamboo forests, providing a mathematical basis for automated identification.

[0089] Logistic regression models output continuous probability values, requiring a threshold to transform them into specific "affected / unaffected" classification decisions. To scientifically determine this threshold, this invention uses receiver operating characteristic (ROC) curves for evaluation. In GEE, a validation sample set independent of the training set is used... The model's performance metrics at different discrimination thresholds were calculated using various methods. The ROC curve, plotted with sensitivity on the vertical axis and 1 - specificity on the horizontal axis, was constructed by systematically traversing all possible classification thresholds. The area under the curve (AUC) was used as an indicator of the model's overall discriminative effectiveness; the closer the AUC value is to 1, the better the model's performance. The optimal classification threshold was determined using the Youden exponent maximization principle, i.e., finding the threshold point that maximizes "sensitivity + specificity - 1," which corresponds to the coordinate closest to the upper left corner of the ROC curve. Using this method, this embodiment determined the optimal classification threshold to be 0.4744, which optimally balances the model's ability to identify disaster-stricken areas with its ability to exclude unaffected areas.

[0090] Combining the trained logistic regression equation with the optimal classification threshold determined by the ROC curve constitutes a complete snow disaster discrimination system for bamboo forests. In practical applications, for any pixel in the target area, the catastrophic difference values ​​of its eight key features are extracted and substituted into the discrimination equation to calculate the y-value. If the threshold is not found, the pixel is determined to be in a snow disaster-affected area; otherwise, it is determined to be in an unaffected area. The entire model training and threshold determination process is implemented through automated scripts in the GEE platform, ensuring the reproducibility and scalability of the method. This rigorous implementation process not only produces a discriminative model with clear physical meaning, but also determines the optimal classification standard through statistical optimization methods, providing a reliable technical solution for high-precision and automated identification of snow disaster-affected areas in large-scale bamboo forests.

[0091] In this embodiment of the invention, the snow disaster discrimination model and its optimal classification threshold are applied to classify the snow disaster status of pixels in all bamboo forest distribution areas within the target region one by one, and output a spatial distribution map of bamboo forest snow disaster-affected areas in the form of a binary image. Specific implementation methods include:

[0092] To achieve accurate assessment of snow disaster-affected areas in bamboo forests, the model's application scope must be limited to the bamboo forest distribution area. This embodiment of the invention uses forest resource distribution data released by the national forestry authorities, or bamboo forest thematic vector data extracted from multi-temporal remote sensing imagery using an object-oriented classification method. In the GEE platform, this vector data is used as a spatial mask to crop the target area, ensuring that all subsequent calculations are performed only on the bamboo forest-covered area. This crucial preprocessing step effectively eliminates interference from other land cover types, providing an accurate spatial analysis basis for precise disaster assessment and ensuring the relevance and reliability of the final results.

[0093] Using the disaster difference imagery of eight key feature variables (masked) as input data, spatial calculations of a logistic regression model are performed on the GEE platform. Specifically, the regression equation, which includes constant terms and feature coefficients, obtained through training, is programmed and embedded into the platform's map algebra module to calculate the predicted value of each pixel within the bamboo forest distribution area in parallel. This process generates a continuous raster layer of predicted values, where the value of each pixel represents the logistic regression output value indicating whether that location belongs to the snow disaster state. The numerical range is determined by the model, reflecting the continuous gradient distribution of the probability of disaster at different locations.

[0094] After obtaining the predicted value raster layer, binary classification processing is performed using the optimal classification threshold determined through ROC curve analysis. The specific discrimination rule is as follows: when the predicted value of a pixel is greater than the set optimal classification threshold, the pixel is marked as a "disaster-affected pixel"; otherwise, it is marked as a "non-disaster-affected pixel." This decision-making process transforms continuous predicted values ​​into discrete classification results, generating a spatially complete binary raster of snow disaster distribution. Different values ​​represent two different states of bamboo forests, clearly defining the spatial differentiation pattern between disaster-affected and non-disaster-affected areas.

[0095] The binary classification results are processed using cartographic techniques to generate thematic maps that conform to professional cartographic standards. By setting an appropriate color scheme, disaster-stricken areas are rendered in striking warning colors, while unaffected areas are rendered in natural green tones. Map elements such as scale bars, legends, and north arrows are added. Simultaneously, complete spatial reference information and metadata descriptions are added to the output data. The final results can be exported in a standard geospatial data format, supporting further analysis, area calculation, and cartographic enhancement in professional geographic information system software. The output "Spatial Distribution Map of Snow Disaster-Stricken Areas in Moso Bamboo Forests" visually demonstrates the spatial pattern and severity of the snow disaster's impact, providing a scientific basis for post-disaster relief deployment, loss assessment, and ecological restoration planning. This embodies the complete technical chain of this invention, from theoretical methods to practical application.

[0096] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for determining the snow disaster area of Phyllostachys edulis forest based on GEE platform, characterized in that, The method comprises the following steps: S1: using the GEE platform and combining the preset cloud coverage threshold and time window, performing radiation calibration and atmospheric correction on the Sentinel-2 image sequence before and after the snow disaster at a specific time phase, and generating a pre-processed multi-temporal remote sensing image set before and after the disaster with physical meaning of surface reflectivity; S2: using the multi-temporal remote sensing image set before and after the disaster, completing multi-dimensional feature extraction covering spectral band reflectivity, multiple vegetation indices and NDVI derived texture features for the two time phases before and after the disaster, and generating a feature initial set containing all features of the pre-disaster state and the post-disaster state; S3: for the feature initial set, by calculating the difference between the post-disaster feature value and the pre-disaster feature value and implementing significance filtering based on statistical hypothesis testing, a subset of disaster feature variables containing only the features that have changed significantly before and after the snow disaster is obtained; S4: around the subset of disaster feature variables, the redundancy elimination based on correlation analysis and the feature importance screening based on multiple machine learning algorithms are sequentially developed, and a key feature variable combination composed of core spectral bands, vegetation indices and texture features is output; S5: using the disaster information represented by the key feature variable combination, performing parameter training of the logistic regression model and evaluating the model discrimination performance with the ROC curve, and establishing a snow disaster discrimination model with a clear mathematical expression and an optimal classification threshold that can optimally distinguish between disaster and non-disaster pixels; S6: applying the snow disaster discrimination model and the optimal classification threshold, classifying the pixels of all bamboo forest distribution areas in the target area one by one according to the snow disaster state, and outputting a bamboo forest snow disaster area spatial distribution map in the form of a binary image.

2. The method according to claim 1, wherein the method is characterized in that, In step S1: The preset time window is the cloud-free image period in the growing season of the vegetation within one year before the snow disaster, and the image period after the cloud and snow influence subsides after the snow disaster; The preset cloud coverage threshold is set to a predetermined percentage threshold for the whole image; Radiation calibration and atmospheric correction are completed using the Sen2Cor algorithm embedded in the GEE platform; The generated multi-temporal remote sensing image set before and after the disaster has consistent spatial resolution and geographic coordinate reference.

3. The method according to claim 1, wherein the method is characterized in that, In step S2: The spectral band reflectivity at least includes B4 red band, B7 red edge 3 band, B8 near infrared band and B11 short wave infrared 1 band of Sentinel-2; The multiple vegetation indices at least include atmospheric resistance vegetation index ARVI, normalized moisture index NDMI, red edge normalized vegetation index NDRE1, normalized water index NDWI and ratio vegetation index RVI; The NDVI derived texture feature is a homogeneity texture feature calculated based on the NDVI image through the gray level co-occurrence matrix; The feature initial set contains all feature variables calculated before and after the two time phases.

4. The method according to claim 1, wherein the method is characterized in that, In step S3: The difference between the post-disaster feature value and the pre-disaster feature value is specifically the difference between the feature value of the post-disaster phase and the feature value of the pre-disaster phase, and the disaster difference value of each feature is obtained; The significance filtering based on statistical hypothesis testing specifically adopts the two-sample t-test method; The t-test is set with a predefined significance level to screen out the feature variables with significant changes; The obtained subset of catastrophic feature variables is the basis for subsequent feature screening process.

5. The method according to claim 1, wherein the method is characterized in that, In step S4, the redundancy elimination based on correlation analysis specifically includes: Calculate the Pearson correlation coefficient between each pair of variables in the subset of catastrophic feature variables; Set a predefined high correlation decision threshold; For the variable pairs with correlation coefficient absolute value exceeding the decision threshold, calculate the average correlation coefficient absolute value of each variable with all other variables in the subset; Remove the variable with higher average correlation coefficient absolute value to retain the features with more independent information and lower redundancy.

6. The method according to claim 5, wherein the method is characterized in that, In step S4, the feature importance screening based on multiple machine learning algorithms specifically includes: Use LASSO regression analysis to shrink the regression coefficients by L1 regularization, compress the coefficients of unimportant variables to zero, and thus achieve feature screening; At the same time, use random forest algorithm to sort the features by calculating the average impurity reduction of each feature in the process of building multiple decision trees; Take the intersection of the variables selected by both LASSO regression analysis and random forest algorithm as the final key feature variable combination; The key feature variable combination consists of eight variables, including the homogeneity texture feature of NDVI, B4 red light band, B7 red edge 3 band, ARVI, NDMI, NDRE1, NDWI, and RVI.

7. The method according to claim 6, wherein the method is characterized in that, The variables in the key feature variable combination have clear physical meaning and calculation formula: The formula for calculating the atmospheric resistance vegetation index ARVI is: ; The normalized moisture index NDMI is calculated according to the formula: ; The calculation formula of the red edge normalized difference vegetation index NDRE1 is: ; The calculation formula of the ratio vegetation index RVI is: ; Where B2, B4, B5, B8, and B11 represent the surface reflectance values of blue light, red light, red edge 1, near-infrared, and short-wave infrared 1 band of Sentinel-2 satellite image, respectively. B2: blue light, 490nm, B4: red light, 665nm, B5: red edge 1, 705nm, B8: near-infrared, 842nm, and B11: short-wave infrared 1, 1610nm.

8. The method according to claim 1, wherein the method is characterized in that, In step S5: The logistic regression model trains parameters by maximum likelihood estimation method, and the discriminant equation established by it is the weighted sum of linear function containing the key feature variable combination and constant term; The ROC curve is used to evaluate the performance of the discriminant equation at different classification thresholds; The optimal classification threshold is determined based on the analysis results of the ROC curve to balance the classification accuracy, sensitivity, and specificity of the model.

9. The method according to claim 1, wherein the method is characterized in that, In step S6: Classify the snow disaster status of each pixel in the target area one by one, and the specific criterion is: calculate the predicted value y of each pixel, if y is greater than the optimal classification threshold, then the pixel is in the disaster state, otherwise it is in the non-disaster state; In the binary image, different pixel values are used to represent the disaster and non-disaster areas respectively; The output snow disaster area spatial distribution map of Phyllostachys edulis forest contains geographic coordinate information, which can be directly loaded and area calculated and spatially analyzed in GIS software.

Images