Tree species identification method, device and equipment based on dense time sequence images

Abstract

The embodiment of the invention provides a tree species identification method, device and equipment based on a dense time sequence image, and the method comprises the steps: obtaining a dense long time sequence multispectral remote sensing image and topographic data of a research area; carrying out image wave band fusion according to the dense long-time-sequence multispectral remote sensing image to construct 12 vegetation spectral indexes; performing land utilization classification on ground features in the dense long-time-sequence multispectral remote sensing image of the research area by utilizing a linear spectrum hybrid model in combination with topographic data, dividing the ground features into buildings, water bodies, other ground features and a forest land data set, and performing pre-classification based on the extracted forest land data set; dividing the forest land data set into an evergreen forest and a deciduous forest by combining an RVI threshold method with a linear spectrum hybrid model; according to the forest land data set, Savitzaky-Golay filtering transformation and spectral differential transformation are carried out, and an NDVI transformation image set is obtained; and performing combination according to the vegetation spectral index NDVI transformation image set as classification features, and discussing the classification effect of the tree species in the research area through a random forest model.

Description

Technical Field

[0001] The present application relates to the technical field of forest resource survey, and in particular to a tree species identification method, device and equipment based on dense time-series images. Background Art

[0002] Tree species information is a key parameter of interest to ecologists and forest managers. Accurately obtaining information on tree species distribution within forest landscapes is crucial for ecological conservation, resource surveys, and environmental monitoring. Tree species composition is an indicator of biodiversity, and information on forest spatial distribution and composition is crucial for biodiversity assessment and monitoring, as well as for the conservation and protection of forest ecosystems. Different tree species have varying growth rates, wood density, and biomass distribution patterns, factors that directly influence their carbon storage capacity. Accurate identification of tree species allows for more precise estimation of regional forest carbon stocks and the abundance of different types of forest resources.

[0003] Traditionally, forest resource inventories have relied on statistical methods and field-based forestland inventory surveys, which are costly, time-consuming, and space-limited, making them difficult to implement in a short period of time. Currently, there are few studies on tree species classification over large areas. Remote sensing data is characterized by large-scale, long-term monitoring. Compared with traditional field work, it can capture forest type composition and forest structure information over large areas and inaccessible areas through multi-band and multi-mode sensors. It has been widely used in fields such as forest vegetation monitoring. A large number of studies have focused on the use of high-resolution or extremely high-resolution imagery, but these datasets are expensive, have low temporal resolution, and are difficult to achieve large-scale tree species identification. In addition, tree species identification in non-pure forest areas is more susceptible to the influence of farmland, buildings, water bodies, and other land features, making classification more difficult. The severe mixing of land features is the main reason for the current low accuracy of tree species identification.

[0004] In addition, the spectral characteristics of different forest tree species are similar, and the identification of tree species is easily affected by the forest stand structure and environmental background. The current research method is prone to the "different objects with the same spectrum" phenomenon when using hyperspectral images for tree species classification. Even if there are phenological differences in the spectral curves of different dominant tree species, these phenological differences are reflected in the phase shift differences of the pixel spectral curves of long-term images. In particular, the overall characteristics of the phenological curves of the dominant tree species in broad-leaved forests are similar, making them difficult to distinguish. How to expand the phenological differences between tree species is still an urgent problem to be solved for tree species identification. Summary of the Invention

[0005] The embodiments of the present application provide a tree species identification method, device and equipment based on dense time-series images, which can perform large-scale tree species identification based on the integration of temporal dynamic features into dense time-series images, expand the phenological differences between different tree species, thereby improving the recognition accuracy of forest tree species and realizing rapid and effective monitoring of tree species diversity.

[0006] To this end, the following technical solutions are provided in the embodiments of the present application:

[0007] In the first aspect, an embodiment of the present application provides a tree species identification method based on dense time-series images, including: obtaining dense long-time-series multispectral remote sensing images and terrain data of the study area; the dense long-time-series multispectral remote sensing images include all available remote sensing forest images with cloud cover less than 20%; performing image band fusion based on the dense long-time-series multispectral remote sensing images to construct a vegetation spectral index; the vegetation spectral index is used to indicate the dynamic change characteristic variables of vegetation growth within the year; using a linear spectral mixture model combined with terrain data to classify the land use of the objects in the dense long-time-series multispectral remote sensing images of the study area into buildings, water bodies and other The method uses the RVI threshold method combined with the linear spectral mixture model to divide the forest dataset into evergreen forest and deciduous forest; the forest dataset is filtered and transformed into NDVI transformed image set; the NDVI transformed image set indicates the time-series-based remote sensing phenological characteristics of plants, and the time-series-based remote sensing phenological characteristics indicate the growth and development status and periodic changes of plants in different seasons; the tree species in the study area are classified by combining the vegetation spectral index and the NDVI transformed image set as classification features through a random forest model.

[0008] In this implementation, traditional forest tree species identification methods rely on ground surveys, which are costly, time-consuming, and have limited spatial coverage. Remote sensing data, on the other hand, allows for efficient monitoring over large areas, particularly in inaccessible areas, allowing for rapid acquisition of large amounts of image data. The use of dense time-series imagery enables continuous monitoring of forest changes, making it suitable for large-scale and high-frequency tree species identification. Combining dense time-series imagery with vegetation indices (such as RVI and NDVI) better reflects plant growth and captures the dynamic characteristics of different tree species across different seasons. These dynamic characteristics can enhance phenological differences between tree species and improve classification accuracy. This solution utilizes multiple data sources (multispectral remote sensing imagery, Sentinel-1 radar data, and terrain data) for tree species identification. Through feature fusion and model training, it comprehensively considers various environmental factors, resulting in more accurate tree species classification, particularly in complex forests and non-pure forest areas. The spectral characteristics of different tree species can be very similar in some cases, particularly for dominant species within broadleaf forests. By incorporating temporal dynamic features and phenological changes, the "same spectrum for different species" problem can be effectively addressed, improving tree species classification accuracy. For example, red edge vegetation indices such as the Modified Red Edge Simple Ratio Index (MSR) can help identify the differences between different tree species in time series and reduce the impact of spectral mixing. This method can be effectively applied to a wide range of areas and is suitable for tree species classification in various forest types and non-pure forest areas (such as farmland, buildings, water bodies, etc.). By combining different remote sensing data, the difficulty of identification caused by mixed objects can be solved and the classification accuracy can be improved. Compared with traditional ground survey methods, this remote sensing technology can significantly save time and costs, while providing real-time, continuous monitoring data to help forest managers and ecologists efficiently obtain large-scale tree species distribution information.

[0009] As a feasible implementation, the image band fusion is performed based on the dense long-time multispectral remote sensing image to construct the vegetation spectral index, including: preprocessing the dense long-time multispectral remote sensing image and terrain data to obtain a forest growth image, and the preprocessing includes image registration, resampling, time series and cloud screening, radiation correction, geometric correction, band synthesis, cloud and shadow masking and image cropping.

[0010] In this embodiment, the preprocessing step is to ensure that the acquired remote sensing data is of high quality and accuracy and is suitable for subsequent tree species identification and forest monitoring tasks. Specifically, the preprocessing steps, especially image registration and resampling, ensure the temporal and spatial consistency of image data from different times and sources, providing a reliable basis for subsequent analysis. Through steps such as cloud screening, radiation correction, and geometric correction, the noise in the image is reduced, the spectral differences of different tree species are enhanced, and the accuracy of tree species identification is improved. Cloud and shadow masks and other preprocessing steps ensure that remote sensing data in complex environments (such as cloud cover, different terrains, etc.) still have good analysis capabilities and are suitable for different types of forest and landform classification. Cropping and other preprocessing operations reduce the amount of calculation and data redundancy, making subsequent analysis more efficient and accurate, especially in forest monitoring tasks in large areas, and can quickly provide accurate analysis results. This preprocessing scheme ensures the quality of remote sensing data through a series of refined steps. From image registration to cropping, each link lays a solid foundation for subsequent forest growth monitoring and tree species identification. Through these preprocessing steps, the accuracy and efficiency of forest monitoring can be effectively improved, making this technical solution widely applicable to large-scale, high-frequency forest resource surveys, ecological protection, and tree species identification.

[0011] As a feasible implementation, the resampling includes: resampling the terrain data to a spatial resolution of 10m using a linear sampling method; and resampling the band of the dense long-time multispectral remote sensing image with a pixel resolution of 20m to 10m using a bicubic interpolation method.

[0012] In this implementation, by resampling terrain data and remote sensing imagery, the spatial resolution of information from different data sources can be kept consistent, facilitating subsequent analysis and data fusion. For example, unifying the resolution of terrain data and remote sensing imagery to 10m can effectively avoid spatial matching issues caused by inconsistent resolutions. Bicubic interpolation and linear interpolation are respectively suitable for resampling different types of data: the former is suitable for refined image processing, while the latter is suitable for simple and efficient processing of terrain data. Combining the two ensures high precision of the resampled data while balancing computational efficiency. For example, through bicubic interpolation, remote sensing imagery can maintain good edge clarity and detailed information even after increasing its spatial resolution, making it suitable for tasks requiring precise monitoring of forest changes. Linear interpolation, when resampling terrain data, preserves the overall elevation trend, ensuring no information loss due to oversmoothing. The resampled high-resolution data can better support a variety of subsequent analytical tasks, such as forest growth and tree species identification. High-resolution terrain data helps to more accurately simulate the impact of terrain on the forest ecosystem, while high-resolution remote sensing imagery can provide more detail, assisting in more accurate forest resource assessments. Resampling methods are simple and efficient, especially linear interpolation and bicubic interpolation. For large-scale forest area monitoring, they not only ensure data quality but also have advantages in time and computation. This resampling technology solution uses a combination of linear interpolation and bicubic interpolation to improve spatial resolution while maintaining the accuracy and detail of terrain and remote sensing image data. This method can not only ensure data quality but also improve the accuracy of subsequent analysis, and is particularly suitable for tasks such as forest monitoring and resource assessment.

[0013] As a feasible implementation, the land use classification of the objects in the dense long-time series multispectral remote sensing images of the study area using the linear spectral mixture model combined with terrain data includes:

[0014] The linear spectral mixture model refers to the spectral reflectance of a pixel in a certain band. It is a linear combination of the reflectance of each component constituting the pixel with its area ratio as the weight coefficient, which is expressed as:

[0015]

[0016] Where: R iλ is the spectral reflectance of the i-th pixel in the λ band; r kλ is the spectral reflectance of the kth basic component in the uth band; f ki is the abundance of the kth endpoint element in the i-th pixel; ξ iλ It is the residual value, representing the multiple reflections and transmissions of light between pixel components, which has a nonlinear mixing effect; n is the number of terminal elements, and m represents the number of basic components that make up the pixel.

[0017] In this embodiment, in remote sensing images, especially in forests, cities and agricultural areas, a single pixel usually contains mixed spectra of multiple land feature types. The linear spectral mixture model can effectively deal with this mixing effect by considering the linear combination of spectral reflectance, and then accurately separate and extract the contribution of different land feature components. Traditional remote sensing image classification methods (such as maximum likelihood method, support vector machine, etc.) usually rely on a single spectral feature, while the linear spectral mixture model can utilize multiple spectral bands of each pixel and combine the proportions of different land feature components to obtain more accurate classification results. By obtaining the abundance information of each component, similar land feature types can be better distinguished and the classification accuracy can be improved. The linear spectral mixture model can be analyzed without strict prior classification information. By inverting the abundance of components in pixels, the model can adaptively identify and classify land feature types, has strong adaptability, and is particularly suitable for large-scale, multi-temporal data sets. For forest areas, the linear spectral mixture model can effectively distinguish between land feature types such as forest cover, bare soil, and water bodies, helping to monitor key indicators such as forest growth and forest cover changes. By identifying the abundance of different components, a more detailed understanding of forest distribution can be achieved, supporting more precise forest resource management. For dense, long-term multispectral remote sensing data, the linear spectral mixture model can process imagery acquired at different time points, helping to analyze spatiotemporal changes in land use. Changes in abundance values ​​can be used to monitor dynamic changes in land use. This model incorporates a residual term, effectively addressing the nonlinear effects caused by object mixing, reducing errors due to spectral mixing and improving reconstruction accuracy. The introduction of the residual term makes the model more adaptable to pixels against complex object backgrounds.

[0018] In a feasible implementation, the vegetation spectral index includes a ratio vegetation index (RVI); and the pre-classification based on the extracted forest dataset to classify the forest dataset into evergreen forest and deciduous forest includes:

[0019] The RVI values ​​of sample points in the deciduous forest in winter and summer from the field survey data were randomly selected for ratio processing, and the minimum value of the relatively stable RVI ratio of the two images was obtained as the separation threshold:

[0020]

[0021] Among them, RVI S is the reference sample of the separation threshold, RVI summer is the RVI value of the summer image pixel, RVI winter is the RVI value of the winter image pixel;

[0022] Obtaining reference sample RVI for deciduous forests s The mean M RVIand standard deviation D RVI ;

[0023] Based on the mean M RVI and standard deviation D RVI and multiple reference sample RVIs s The frequency histogram is drawn to obtain the reference sample RVI that obeys the normal distribution. S Frequency and RVI S Normal distribution curve;

[0024] According to RVI S Frequency and RVI S The normal distribution curve sets a discrimination rule, and deciduous forests and evergreen forests are determined by the discrimination rule.

[0025] In this embodiment, the technical solution effectively utilizes seasonal differences by utilizing the different spectral characteristics of deciduous forests in winter and summer. The spectral reflectance characteristics of vegetation in different seasons are significantly different, especially the RVI values ​​of deciduous forests in winter and summer change significantly, and this change can be used to accurately distinguish different types of woodlands. The method automatically sets the discrimination rules based on statistical analysis and normal distribution curves, making the classification process more objective, flexible and adaptive. Setting the discrimination rules by the mean and standard deviation can cope with changes and noise in sample data, thereby improving classification accuracy. The technical solution is based on the vegetation spectral index RVI, uses seasonal changes to classify woodlands, and sets automatic discrimination rules through statistical analysis methods, achieving efficient distinction between evergreen forests and deciduous forests. Its advantages are that it makes full use of seasonal differences, improves classification accuracy, and has a high degree of automation and adaptability, and can quickly process large-scale remote sensing data.

[0026] As a feasible implementation, the vegetation spectral index includes time series data of the normalized vegetation index NDVI; the filtering transformation and spectral differential transformation are performed on the forest data set to obtain an NDVI transformed image set, including: based on the phenological information of the dominant tree species in the study area, the NDVI time series data is subjected to polynomial filtering transformation using the Savitaky-Golay algorithm to obtain a smoothed and denoised NDVI time series spectral curve of the dominant tree species; the phenological information includes seasonal characteristic information of the dominant tree species, and the seasonal characteristic information indicates the canopy change information of the leaf bud opening period, leaf expansion period, growth period, leaf full color period and leaf fall end period of each tree species; based on the first-order differential transformation and the second-order differential transformation, the NDVI time series spectral curve of the dominant tree species after extracting different spectral parameters is obtained.

[0027] In this embodiment, the Savitzky-Golay filtering algorithm effectively denoises and smoothes the time series data, reducing the impact of noise caused by external factors (such as climate change, observation errors, etc.), so that the data can more accurately reflect the actual growth status of vegetation. Introducing tree species phenological information into NDVI time series data analysis can provide a time reference based on the tree species growth cycle. This information helps to improve the accuracy of time series data analysis, especially to accurately analyze and distinguish the growth stages of different tree species. Through first-order and second-order differential transformations, the growth changes of vegetation can be deeply analyzed, including growth rate and acceleration of change. This not only helps to reveal subtle changes in the growth process of vegetation, but also provides support for subsequent vegetation growth trend prediction and ecological monitoring. By extracting NDVI time series curves of different tree species, the growth characteristics of different tree species in the same area can be effectively distinguished. Especially in complex ecological environments, the growth differences between tree species can be more clearly revealed through this solution.

[0028] As an achievable implementation, the polynomial filtering transformation of NDVI time series data using the Savitaky-Golay algorithm includes:

[0029] For NDVI time series data Y(t), (t=0,1…,Y N-1 ) where N is a positive integer, and for its adjacent 2k+1 data points (where k is a positive integer), fit an l-order polynomial in the local area:

[0030] σ(t)=a0+a1t+a2t 2 +…+a l t l

[0031] Where σ(t) is the fitting polynomial; a l (l=0, 1, ..., l) are coefficients of the polynomial;

[0032] Obtaining polynomial coefficient a based on local polynomial least squares fitting l , so that the square error between the fitting polynomial and the original data in the local area is minimized:

[0033]

[0034] Where E is the square error, C j is the weighting coefficient, Y(t+j) is the actual observation value at time point t+j, and σ(t+j) is the actual fitting value at time point t+j calculated by polynomial fitting;

[0035] Through the above polynomial fitting process, the smoothed data point Y can be obtained (t) , these smoothed data points Y(t) That is, the data after filtering and denoising is used to obtain the NDVI time series spectral curve of the dominant tree species after SG filtering transformation.

[0036] In this embodiment, the smoothed data is more accurate when performing subsequent analysis and modeling. For example, when using NDVI data to construct a vegetation growth model or predict ecological and environmental changes, the denoised data can provide more reliable input and reduce errors caused by data noise. Specifically, the Savitky-Golay algorithm provides powerful denoising and smoothing capabilities for NDVI time series data by minimizing square errors through polynomial fitting and least squares method. This technology not only retains the main trends of the vegetation growth cycle, but also effectively removes high-frequency noise in the data, greatly improving the accuracy and analyzability of the data. Its flexibility, adaptability and efficient calculation give it significant advantages in remote sensing data processing, vegetation monitoring and ecological and environmental research.

[0037] In a feasible implementation, the NDVI time series spectral curve of the dominant tree species after extracting different spectral parameters based on the first-order differential transformation and the second-order differential transformation includes:

[0038] The calculation formulas for first-order differential transform and second-order differential transform are as follows:

[0039]

[0040] In the formula, R(λ i ) is the spectral reflectance value at band i, R(λ i )′,R(λ i )″ are the first-order and second-order values ​​of the spectrum between bands i and i+1, and Δλ is the step size between adjacent bands.

[0041] In this embodiment, by performing first-order and second-order differential transformations on the NDVI time series spectral data, different spectral change characteristics can be extracted, which is of great significance for the identification and classification of dominant tree species. The first-order differential mainly reflects the rapid changes between bands. For example, in the vegetation index (NDVI) curve, the first-order differential can reveal the mutation point of vegetation growth, such as the change in the rapid growth stage or the decline stage. The second-order differential can reflect the smoothness and curvature of the spectral curve. When analyzing vegetation growth, the second-order differential can effectively distinguish the changing trends of different vegetation types, especially in the slow growth or stable period, and can better capture the detailed changes, especially in the period when the seasonal changes are relatively mild. Through the first-order and second-order differential transformations, the sensitivity to vegetation growth changes can be improved, especially the subtle changes in vegetation indices (such as NDVI) can be captured in detail, and the growth status of tree species can be identified. For example, the first-order differential transformation can reveal the rapid changes in vegetation recovery in spring or decline in autumn, while the second-order differential transformation can capture more subtle changes, such as the curvature change of vegetation growth. The NDVI time-series spectral curves of different tree species exhibit distinct patterns of change during their growth. First- and second-order differential transforms can extract richer spectral features, such as the changing trends in vegetation reflectance and the fluctuations in the spectral curves. This information can be used to further differentiate the spectral curves of different tree species, supporting remote sensing identification and monitoring of tree species. First- and second-order differential processing of spectral data can yield more characteristic information, helping to improve the accuracy of remote sensing data analysis. In studies such as vegetation growth cycle analysis, forest health monitoring, and soil moisture analysis, differential transforms can help analysts more precisely identify change points and key features, resulting in more accurate results. Because differential transforms essentially measure the rate of change, they can effectively reduce the impact of low-frequency noise in the raw data. This is particularly true for components with smaller reflectance changes. Differential transforms can amplify these small changes while suppressing insignificant low-frequency fluctuations, thereby reducing noise interference.

[0042] Spectral data processing methods based on first- and second-order differential transforms can accurately reveal important information such as vegetation growth status, tree species characteristics, and environmental changes by extracting the rates of change between different bands. This method has significant advantages in remote sensing data processing, effectively improving the accuracy of data analysis and enhancing the monitoring capabilities of vegetation growth and ecological environments.

[0043] In one achievable embodiment, the method further includes: verifying the accuracy of the tree species identification result, including:

[0044] The tree species identification results were evaluated from the perspectives of overall accuracy and kappa coefficient;

[0045] The overall accuracy evaluation is to compare the pixel-level tree species recognition results of long-term remote sensing image data with the ground truth to evaluate the accuracy of the classification results. The calculation formula for the overall accuracy is as follows:

[0046] Overall accuracy = (TP+TN) / (TP+TN+FP+FN)

[0047] TP represents the number of samples that are truly positive and correctly predicted as positive, TN represents the number of samples that are truly negative and correctly predicted as negative, FP represents the number of samples that are truly negative but incorrectly predicted as positive, and FN represents the number of samples that are truly positive but incorrectly predicted as negative.

[0048] The Kappa coefficient is a statistical indicator for evaluating the consistency between classifiers or evaluators.

[0049] The calculation formula of Kappa coefficient is as follows:

[0050]

[0051] Among them, Po is the sum of the number of correctly classified samples in each category divided by the total number of samples, that is, the overall classification accuracy; assuming that the number of true samples in each category is a1, a2, ..., a C , and the predicted number of samples of each class are b1, b2, ..., b C The total number of samples is n, then:

[0052]

[0053] The calculated result of kappa is -1 to 1, but usually kappa falls between 0 and 1. It can be divided into five groups to represent different levels of consistency: 0.0 to 0.20 is very low consistency (slight), 0.20 to 0.40 is general consistency (fair), 0.40 to 0.60 is moderate consistency (moderate), 0.60 to 0.80 is high consistency (substantial) and 0.80 to 1 is almost perfect.

[0054] In this implementation, the overall accuracy provides an intuitive percentage, indicating the proportion of correctly classified samples to all samples, while the Kappa coefficient takes into account the consistency between the classifier and the actual situation, and can statistically evaluate the performance of the classifier, avoiding misleading due to uneven data distribution. The two indicators of overall accuracy and Kappa coefficient are used to evaluate the accuracy of tree species identification results, thereby ensuring the reliability and effectiveness of remote sensing image data classification. Through the comprehensive evaluation of overall accuracy and Kappa coefficient, this technical solution provides a comprehensive and quantitative accuracy verification method for tree species identification results. This method can not only accurately evaluate the performance of the classifier, but also quantify the accuracy and consistency of the classification results, which is of great significance for tree species identification in multi-time series remote sensing image analysis. In addition, it provides a clear direction and basis for subsequent model optimization and data analysis, and has strong applicability and practical value.

[0055] In the second aspect, the embodiment of the present application also provides a tree species identification device based on dense time-series images, including: an acquisition module, the acquisition module is used to acquire dense long-time-series multispectral remote sensing images and terrain data of the study area; the dense long-time-series multispectral remote sensing images include all available remote sensing forest images with cloud cover less than 20%; a processing module is used to perform image band fusion based on the dense long-time-series multispectral remote sensing images to construct a vegetation index; the vegetation spectral index is used to indicate the dynamic change characteristic variables of vegetation growth within the year; a linear spectral mixture model is used to classify the land use of the ground objects in the dense long-time-series multispectral remote sensing images of the study area into buildings, water, and trees. The extracted forest dataset is pre-classified, and the forest dataset is divided into evergreen forest and deciduous forest by using the RVI threshold method combined with the linear spectral mixture model; Savitzaky-Golay filter transformation and spectral differential transformation are performed on the forest dataset to obtain the NDVI transformed image set; the NDVI transformed image set indicates the time-series-based remote sensing phenological characteristics of plants, and the time-series-based remote sensing phenological characteristics indicate the growth and development status and periodic changes of plants in different seasons; the tree species in the study area are classified by the random forest model based on the vegetation spectral index and the NDVI transformed image set.

[0056] In a third aspect, an embodiment of the present application further provides a computing device comprising at least one memory for storing programs; and at least one processor for executing the programs stored in the memory; wherein the memory is coupled to the processor, and when the program stored in the memory is executed, the processor is used to execute the method described in the first aspect or any possible implementation of the first aspect.

[0057] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, which stores computer instructions. When the computer instructions in the computer-readable storage medium are executed by a computing device, the computing device executes the method involved in the first aspect and its possible implementation methods.

[0058] In a fifth aspect, an embodiment of the present application further provides a computer program product, which includes computer instructions. When the computer instructions are executed by a computing device, the computing device executes the method described in the first aspect and its possible implementation methods.

[0059] It can be understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here.

[0060] The technical solution of this application achieves high-precision tree species identification by integrating multispectral remote sensing images, Sentinel-1 radar data, terrain data, and temporal variation characteristics, combined with machine learning models (such as random forests). The core advantage of this method lies in its comprehensive utilization of multi-dimensional data, which can accurately capture the temporal variation characteristics of vegetation. For example, large-scale tree species identification based on the integration of temporal dynamic characteristics into dense temporal images can expand the phenological differences between tree species, thereby improving the recognition accuracy of forest tree species and realizing rapid and effective monitoring of tree species diversity. At the same time, tree species classification is carried out in an automated manner, which is suitable for remote sensing monitoring and vegetation resource assessment in large areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A flowchart of a tree species identification method based on dense time-series images provided in an embodiment of the present application;

[0062] Figure 2 shows the RVI based on the reference sample s Schematic diagram of the discrimination rules between evergreen forest and deciduous forest based on frequency histogram;

[0063] Figure 3 The seasonal characteristics of dominant tree species in the study area are shown;

[0064] Figure 4 The NDVI time series spectrum curve of dominant tree species after SG filtering transformation is shown;

[0065] Figure 5 The NDVI time series spectrum curves of dominant tree species after SG filtering and differential transformation are shown;

[0066] Figure 6The radar charts of the recognition accuracy of each dominant tree species based on different feature combinations are shown (SI: spectral index; SG: SG filter transform; FD: first-order differential transform; SD: second-order differential transform);

[0067] Figure 7 Flowchart of the application scenario provided in Example 1 of this application:

[0068] Figure 8 A tree species identification device based on dense time-series images provided by an embodiment of the present application is shown;

[0069] Figure 9 A computing device provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0071] In the description of the embodiments of the present application, words such as "exemplary," "for example," or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary," "for example," or "for example" in the embodiments of the present application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary," "for example," or "for example" is intended to present the relevant concepts in a concrete manner.

[0072] In the description of the embodiments of this application, the term "and / or" is simply a description of the association relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent the following three situations: A exists alone, B exists alone, and A and B exist at the same time. In addition, unless otherwise specified, the term "plurality" means two or more. For example, "multiple systems" refers to two or more systems, and "multiple terminals" refers to two or more terminals.

[0073] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly identifying the technical features being referred to. Thus, features specified as "first" or "second" may explicitly or implicitly include one or more of such features. The terms "include," "comprising," "having," and their variations all mean "including but not limited to," unless otherwise specifically emphasized.

[0074] In the description of the embodiments of the present application, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it can be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict.

[0075] In the description of the embodiments of the present application, the terms "first\second\third, etc." or module A, module B, module C, etc. are only used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that the specific order or sequence can be interchanged where permitted so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0076] In the description of the embodiments of the present application, the numbers representing the steps, such as S101, S102, etc., do not necessarily mean that the steps must be executed in this manner. If permitted, the order of the previous and next steps can be interchanged, or they can be executed simultaneously.

[0077] Quantitative information on forest tree species is crucial for achieving sustainable development and protecting the ecological environment. Timely and accurate mapping of forest types is a key component of forest resource inventories. Furthermore, tree species classification and mapping are crucial for calculating terrestrial vegetation carbon stocks and for carbon trading. Forests play a key role in reducing carbon emissions and serve as a significant carbon sink. Methods for estimating regional carbon stocks by establishing species-specific individual tree biomass models based on factors such as tree diameter or standing volume are essential. Accurate and detailed tree species classification and mapping are essential for establishing individual tree diameter at breast height biomass models for major tree species.

[0078] Currently, only 12% of primary forests are protected worldwide, and many more remain at serious risk of degradation. Research indicates that human activities are influencing global forest structure, but it remains unclear whether global human pressures have significantly extended to protected forests and primary forests, which are typically considered less susceptible to human impact. For example, mining significantly damages native forest vegetation, while reclamation of coal mining areas plays a vital role in rebuilding carbon reservoirs and improving ecosystems. Clarifying forest type and species information provides a reliable foundation for coordinated regional planning for the sustainable conservation and utilization of forest resources.

[0079] By achieving rapid and effective monitoring of tree species diversity, biodiversity conservation and research as well as sustainable forest management can be promoted. The level of forest biodiversity plays an important role in maintaining and exerting the functions of forest ecosystems, and is a key factor affecting the stability, productivity and carbon storage of forest ecosystems. Rapid and effective monitoring of tree species diversity is an important part of forest biodiversity conservation, and is of great significance to promoting biodiversity conservation and research as well as sustainable forest management. Moreover, clarifying the spatial information of tree species composition over large areas can help to better understand tree species ecology, such as community dynamics and the contribution of species to ecosystem functions and services. Other environmental studies, such as those focusing on wildlife habitat mapping or estimating insect abundance in forests, also benefit from tree species information. In addition, knowledge of tree species distribution may also affect forest harvesting and government management policies.

[0080] However, existing methods for identifying dominant tree species in regional forests have the following main problems:

[0081] (1) Existing research on tree species extraction mainly relies on high-spectral and high-spatial resolution remote sensing data sources. However, these data are expensive and costly and cannot meet the requirements of large-scale, high-precision tree species identification and mapping.

[0082] (2) When extracting large-scale tree species based on time series images, due to the limitation of image spatial resolution and the widespread existence of pixel mixing during the extraction process, the classification process is easily affected by the interference of different forest types and surrounding background information, resulting in unsatisfactory classification results;

[0083] (3) Most current vegetation remote sensing classification and mapping methods rely on field data from a single time phase. However, field sampling often requires a lot of manpower and material resources. For long-term vegetation classification and dynamic mapping research, reliable historical samples are required. Currently available vegetation classification label samples cannot meet the needs of tree species classification.

[0084] (4) Classify according to the differences in the phenological trajectory characteristics of each tree species. The overall spectral characteristics of vegetation have a high degree of similarity, but dense time series data can accurately reflect the phenological information of each tree species (branching, leafing, flowering and leaf falling, etc.). However, there is currently a lack of effective spectral transformation methods to highlight the differences in phenological trajectories;

[0085] The above problems make it difficult for existing tree species methods in natural forest areas to adapt to the actual needs of regional forestry resource surveys and monitoring.

[0086] In summary, detailed tree species classification is fundamental to forest management planning and disturbance monitoring. Accurate tree species classification provides important basic information and a scientific basis for local forest resource management. It is important to note that image features are crucial for extracting image information, primarily including spectral, spatial, and temporal characteristics. The spatial frequency of an image represents the rate of change of different details within the image, or the degree of variation in grayscale values ​​within the image. Different tree species have complex and variable structures and backgrounds, and texture features lack sufficient stability. Therefore, tree species identification based solely on single-phase images results in low extraction accuracy.

[0087] Based on this, expanding the phenological differences between tree species and exploring the dynamic difference characteristics of canopy spectral phenology of different tree species under complex backgrounds are the keys to identifying tree species. Therefore, this application makes full use of the spectral information, terrain data and radar data of multispectral remote sensing, explores the influence of spectral filtering and differential changes on the results of regional forest tree species identification, and proposes a tree species identification method, device and equipment based on dense time series images. Specifically, the present application provides a tree species identification method, device and equipment based on dense time series images, which is based on dense time series images and can effectively improve the accuracy of tree species identification by combining multiple remote sensing data sources (multispectral images, Sentinel-1 radar data, terrain data, etc.) and multiple data processing methods (such as vegetation index, NDVI transformation, linear spectral mixture model, etc.), especially in large-scale and diverse forest areas. By using temporal dynamic characteristics to expand the phenological differences between tree species, the problem of "different species with the same spectrum" is solved, with strong adaptability and higher efficiency and accuracy than traditional methods.

[0088] Figure 1 Schematic diagram of the process flow of the tree species identification method based on dense time series images provided in the embodiment of the present application (refer to the process framework diagram at the same time Figure 7 ).like Figure 1 The specific steps are as follows:

[0089] S101. Acquire dense long-term multispectral remote sensing images and terrain data of the study area; dense long-term multispectral remote sensing images include all available remote sensing forest images with cloud cover less than 20%.

[0090] Among them, dense long-time multispectral remote sensing images include all bands of the original images, which can be obtained by downloading all available remote sensing satellite images of the study area in the corresponding years (for example, 2019-2022) from the remote sensing satellite image download platform.

[0091] For example, remote sensing satellite image download platforms include: Google Earth Engine (GEE) platform (https: / / code.earthengine.google.com / ), United States Geological Survey (https: / / www.usgs.gov / ), European Space Agency (https: / / scihub.copernicus.eu / ), Geospatial Data Cloud (http: / / www.gscloud.cn / ), etc., and remote sensing satellite images include: Landsat TM / ETM+ / OLI, Sentinel-2MSI, etc.

[0092] In this embodiment, for dense long-term multispectral remote sensing images, including the available bands of the original images, specifically include: using blue (B2: 490nm), green (B3: 560nm), red (B4: 665nm), near infrared (B8: 842nm) with a spatial resolution of 10m and the red edge area (B5: 705nm; B6: 740nm; B7: 783nm) with a spatial resolution of 20m; narrow NIR (near infrared) band (B8a: 865nm); SWIR (short wave infrared) band (B11: 1610nm; B12: 2190nm).

[0093] The terrain data used is the NASADEM (NASA Digital Elevation Model) dataset. NASADEM data is a global digital elevation model (DEM) jointly released by the National Aeronautics and Space Administration (NASA) and the United States Geological Survey (USGS). It is generated based on NASA's satellite data (such as SRTM - Shuttle Radar Topography Mission) and other satellite measurement data to provide terrain height information on a global scale. Specifically, the NASADEM dataset is a reprocessing of the digital elevation model data generated by the Shuttle Radar Topography Mission (SRTM) data, by incorporating data mainly from the Ice, Cloud and Land Elevation Satellite (ICESat), the Earth Science Laser Altimeter System (GLAS) and the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) instruments.

[0094] In some embodiments, in this step, forest resource data from other sources in the study area are also obtained, including Sentinel-1 radar data, forest inventory data, and field survey data.

[0095] Sentinel-1 radar data corresponds to Sentinel-1 data with a spatial resolution of 10m for each year. Forest inventory data can be recent forest land change survey data provided by the local forestry bureau and approved by the local forestry bureau. Forest inventory data primarily uses forest plots as survey units, including tree species composition, age structure, tree density, canopy density, height, diameter at breast height, etc. Forest inventory data is used to gain a preliminary understanding of the vegetation types and distribution in the study area, providing a more accurate survey scope for subsequent field surveys. Field survey data is obtained through field surveys conducted under the guidance of technical personnel from the local forestry bureau in areas where tree species are unclear or boundaries are vague in existing forest inventory data.

[0096] For example, for complex areas where field surveys are difficult to conduct and areas where network signal sampling is difficult due to poor quality, drones equipped with multispectral sensors can be used for aerial photography-assisted sampling under the guidance of technical personnel from the local Forestry Bureau to obtain field survey data. At the same time, several representative dominant tree species forests (Larix principis-rupprechtii, Pinus tabulaeformis, Populus davidiana, Quercus liaotungensis and other tree species) are selected for different dominant tree species and divided into sampling areas (more than 40 for each dominant tree species). At the same time, it is necessary to comprehensively consider the differences in forest age structure (referring to the age distribution of trees), understory cover (i.e., the degree to which the ground is covered by plants) and stand density (the number of trees per unit area), and meet the spatial range of each forest training sample area. The standard setting of at least 30*30m area is met. In addition, a random starting point systematic sampling strategy is adopted. Trees with a breast diameter greater than 4m are measured in diameter and height and the tree species are recorded. At the same time, a laser radar (Simultaneous Localization and Mapping, SLAM) device is used to scan the sample plot to accurately extract tree information and provide strong support for tree species identification.

[0097] S102. Construct a vegetation spectral index by performing image band fusion based on dense long-term multispectral remote sensing images; the vegetation spectral index is used to indicate characteristic variables of dynamic changes in vegetation growth within the year.

[0098] In some implementations, step S102 may be implemented by the following steps:

[0099] S1021, preprocessing dense long-term multispectral remote sensing images and terrain data to obtain forest growth images.

[0100] Among them, the preprocessing of dense long-term multispectral remote sensing images includes: time series and cloud screening, radiation correction, geometric correction, band synthesis, cloud and shadow masks and image cropping.

[0101] Cloud screening of dense, long-term multispectral remote sensing imagery is performed to remove cloud cover, as clouds can affect the quality of ground observations and lead to inaccurate data. Cloud detection algorithms (such as the Fmask algorithm) are typically used to detect and remove clouds and their shadows. Time series selection helps obtain a more comprehensive picture of vegetation dynamics. Radiometric correction converts the raw radiometric values ​​of dense, long-term multispectral remote sensing imagery into usable ground reflectance data. Raw data collected from dense, long-term multispectral remote sensing imagery is affected by factors such as atmospheric conditions and sensor performance, making radiometric correction essential for accurate analysis. Geometric correction is the process of spatially aligning images to ensure they match the geographic coordinate system and eliminate spatial errors caused by factors such as Earth curvature, sensor angle, and orbital variations. Data from different bands are synthesized, typically selecting bands of interest based on research needs. The Sentinel-2 satellite has 13 bands, some of which have different spatial resolutions. For example, the pixel resolution of bands B5, B6, B7, B8A, B11, and B12 is 20 meters, while the resolution of bands B2, B3, B4, and B8 is 10 meters. Using the cloud detection results from the imagery, areas of cloud and cloud shadows are removed from the imagery. This is crucial for subsequent data analysis, as clouds and shadows can affect the true reflectance of ground objects. Cropping remote sensing images according to the boundaries of the study area, removing irrelevant areas and retaining only the areas of interest, reduces computational effort and improves processing efficiency.

[0102] It should be noted that dense long-term multispectral remote sensing images (hereinafter referred to as satellite remote sensing images) are time series images. Since different bands of Sentinel-2 have different resolutions, in order to unify the data format, it is usually necessary to resample the 20m resolution bands (such as B5, B6, etc.) to 10m resolution. Specifically, the pixel resolution of the B5, B6, B7, B8A, B11, and B12 bands of Sentinel-2 is 20m. In this embodiment, the bicubic interpolation method is used to resample the bands to 10m. The bicubic interpolation method has an edge enhancement effect and can better maintain the fine structure of the image, but the amount of calculation is large. Using the GEE platform for calculation can well solve the weaknesses of the method. Next, the B2, B3, B4, and B8 bands and their resampled bands are reprojected to the EPSG:4326 coordinate system and the bands are fused into a new forest growth image.

[0103] For example, the pre-processing platform for satellite remote sensing images may be Google Earth Engine platform, ENVI software, Sen2Cor software, ArcGIS software, SNAP software, etc.

[0104] Exemplarily, the NASADEM dataset is preprocessed, including image registration, radiation correction, geometric correction, resampling, image stitching and cropping. For example, the NASADEM dataset, which is terrain data, is resampled to a spatial resolution of 10m using a linear sampling method, and its slope and aspect are calculated in ArcGIS. By resampling the NASADEM data to a resolution of 10 meters, more detailed terrain information can be provided, which is suitable for high-precision analysis. By calculating the slope and aspect, the terrain features are further quantitatively analyzed, which facilitates the land use classification of the features in the dense long-term multispectral remote sensing images of the study area in the subsequent steps, thereby improving the accuracy of the classification.

[0105] It should be noted that slope refers to the degree of inclination of the terrain, usually expressed as an angle. Slope is an important terrain parameter in environmental studies, land use planning, and landslide prediction. In ArcGIS, you can use the "Slope" tool to calculate slope. This tool calculates the slope value of each pixel based on the elevation value of each pixel in the digital elevation model (DEM) by analyzing the elevation changes of surrounding pixels. Neighborhood-based calculation methods, such as the 8-neighbor method, are commonly used. Aspect indicates the horizontal direction of the terrain surface and is usually expressed in degrees. Aspect indicates the horizontal direction of the terrain surface and is usually expressed in degrees. In ArcGIS, you can use the "Aspect" tool to calculate aspect. This tool analyzes the relative elevation information of each pixel to determine its horizontal direction (0° represents north, 90° represents east, 180° represents south, and 270° represents west). Slope and aspect provide additional terrain information for tree species identification, helping to improve the classification accuracy and reliability of the model. These terrain features can reflect the growth environment of trees. Combined with the spectral data of remote sensing images, we can better understand the spatial distribution and growth patterns of different tree species, and provide more accurate support for tree species identification, ecological assessment, forest management and other tasks.

[0106] For example, preprocessing for aerial photography captured by drones equipped with multispectral sensors includes image registration, radiometric correction, geometric correction, image stitching, and cropping, helping to eliminate the effects of different times and locations and ensure data consistency. If the processing level is Collection2Level-2, it indicates that the product has already undergone basic preprocessing, including geometric correction, radiometric correction, and atmospheric correction, and no further processing is required.

[0107] S1022: performing image band calculation based on information of different bands of the forest growth image to obtain a vegetation spectral index table of the tree species.

[0108] In this step, the vegetation indices include: normalized vegetation index, improved red-edge normalized vegetation index, ratio vegetation index, red-edge chlorophyll index, enhanced vegetation index, vegetation attenuation index, greenness index, improved red-edge ratio vegetation index, new inverted red-edge chlorophyll index, modified soil-adjusted vegetation index, ground chlorophyll index, chlorophyll absorption index, etc., as shown in Table 1.

[0109] Table 1 Spectral indices based on Sentinel-2 remote sensing images (12)

[0110]

[0111] As shown in Table 1, the above vegetation indices participate in subsequent classification as bands, which can effectively reflect the dynamic change characteristics of crop growth within the year. Specifically, Normalized Difference Vegetation Index (NDVI): It is widely used to reflect the density and health of vegetation through the ratio of red light and near-infrared bands. Improved Red-Edge Normalized Vegetation Index (IRECI): It uses the red edge band (the transition zone between red light and near-infrared) to more accurately describe the growth status of vegetation. Ratio Vegetation Index (RVI): It uses the ratio of red light and near-infrared bands to reflect the growth of vegetation. Red-edge Chlorophyll Index (CRECI) red-edge): The vegetation index calculated in the red-edge band is better adapted to the spectral characteristics of different plants. Enhanced Vegetation Index (EVI): For dense vegetation areas, EVI has stronger adaptability than NDVI and can better reduce the impact of the atmosphere. Vegetation Senescence Reflectance Index (PSRI): Evaluates the decline of vegetation through the ratio of the red and green light bands. Green Leaf Index (GLI): Reflects the greenness of plants and is suitable for analyzing the growth status of plants. Soil Adjusted Vegetation Index (SAVI): Based on NDVI, a soil adjustment coefficient is added to eliminate the influence of soil background. These vegetation indices can reflect the dynamic changes of crops and vegetation under different times and environmental conditions, and are especially suitable for monitoring and analyzing vegetation growth conditions and land cover types. The Modified Red Edge Simple Ratio Index (MSR) is a vegetation index based on the red edge band and is often used to enhance sensitivity to vegetation chlorophyll content, nitrogen status, or stress. The Meris terrestrial chlorophyll index (MTCI) estimates the chlorophyll content of vegetation canopies. This index is sensitive to moderate to high chlorophyll concentrations and is suitable for agricultural, ecological, and forest health monitoring. The Modified Normalized Difference Vegetation Index with Red Edge (mNDVI) is an optimized vegetation index that incorporates the red edge band. It is designed to enhance sensitivity to vegetation chlorophyll content, nitrogen status, and stress, while reducing the saturation problem of traditional NDVI in high biomass areas. The Modified Chlorophyll Absorption in Reflectance Index 2 (MCARI2) is an improved vegetation index that enhances sensitivity to chlorophyll content while reducing the influence of soil background and canopy structure. It is an optimized version of the original MCARI and is widely used in agricultural remote sensing, vegetation physiological status monitoring, and other fields. By calculating these vegetation indices and inputting them as feature variables into subsequent classification models, classification accuracy can be effectively improved, especially in the accurate identification and analysis of forest tree species.

[0112] S103. Use the linear spectral mixture model combined with terrain data to classify the land use of features in the dense long-term multispectral remote sensing images of the study area into buildings, water bodies and other features, and forest datasets. Pre-classify the forest dataset based on the extracted forest dataset and divide it into evergreen forest and deciduous forest.

[0113] In this step, the Linear Spectral Mixture Model (LSMM) assumes that the spectral reflectance of each pixel in the mixed spectrum is a combination of multiple endmembers (basic components) within the pixel and their abundances (weights). Specifically, selecting appropriate forest endmember spectra can significantly improve the accuracy of spectral decomposition. Linear spectral decomposition analysis can be applied to all multispectral data, but the LSMM requires that the number of endmembers must be less than or equal to the number of sensor bands plus one.

[0114] For example, the linear spectral mixture model refers to the spectral reflectance of a pixel in a certain band, which is a linear combination of the reflectances of the components constituting the pixel with their area ratios as weight coefficients, and is expressed as:

[0115]

[0116] Where: R iλ is the spectral reflectance of the i-th pixel in the λ band (in other words, it represents the color of a pixel in the image at a specific band); r kλ is the spectral reflectance of the kth basic component in the uth band; f ki is the abundance of the kth end-member in the ith pixel (or the proportion of the end-member in the pixel, with abundance ranging from 0 to 1, indicating the relative proportion of the end-member in the pixel); ξ iλ It is the residual value, representing the multiple reflections and transmissions of light between pixel components, and has a nonlinear mixing effect (this residual term is mainly used to compensate for the multiple reflections and transmissions of light within the pixel, which causes some nonlinear changes in the spectral characteristics of the image); n is the number of terminal elements, and m represents the number of basic components that make up the pixel.

[0117] Specifically, the linear spectral mixture model treats the spectral reflectance of each pixel as a linear combination of multiple land features (such as water bodies, vegetation, buildings, etc.), relying on the spectral characteristics of these components and their relative abundance in the pixel. Endmembers: These are the spectral reflectances representing different land feature types. For example, vegetation, water bodies, bare soil, etc. can all be used as endmembers. Abundance: refers to the relative proportion of each endmember in the pixel, which is usually estimated by optimization algorithms (such as least squares, non-negative matrix decomposition, etc.). Residual: Due to the nonlinear mixing effect of multiple components in the pixel, the model takes into account the residual value to better fit the spectral data. The goal of this model is to perform land use classification by inverting the abundance of each component in each pixel. Through the linear spectral mixture model, combined with terrain data, the remote sensing image is decomposed into data sets such as buildings, water bodies, and woodlands.

[0118] It's worth noting that in this step, the extracted forest dataset can be pre-classified using a hierarchical classification strategy, dividing it into evergreen and deciduous forests. This strategy involves first classifying the vegetation into evergreen and deciduous forests, and then identifying each tree species within each. For example, the vegetation can be first classified into evergreen and deciduous forests, and then Larix principis-rupprechtii, Pinus tabulaeformis, Populus davidiana, Quercus liaotungensis, and other tree species can be extracted from the evergreen forests to achieve hierarchical classification.

[0119] In some embodiments, deciduous forests and evergreen forests can be distinguished based on the ratio vegetation index (RVI). RVI is a vegetation index designed based on the spectral reflectance characteristics of plants in remote sensing images. The ratio vegetation index can provide important information about vegetation reflectance and is an important indicator for testing vegetation abundance. It is particularly sensitive to vegetation in high-density coverage conditions. It describes the growth status of ground vegetation by calculating the reflectance ratio of specific bands. RVI is generally calculated using the following formula shown in Table 1:

[0120] RVI=(NIR-Red) / (NIR+Red)

[0121] Where NIR is the reflectance in the near-infrared band, and Red is the reflectance in the red band. For vegetation, a higher RVI value generally indicates greater vegetation cover, while a lower RVI value indicates less vegetation cover. By adjusting the threshold multiple times and referring to previous research, we can determine a threshold range that effectively extracts gradient variations in evergreen forests.

[0122] For example, the forest dataset can be divided into evergreen forest and deciduous forest by the following steps:

[0123] S1031. Randomly select the RVI values ​​of sample points of the deciduous forest images in winter and summer from the field survey data, perform ratio processing, and obtain the minimum value of the relatively stable RVI ratio of the two images as the separation threshold:

[0124]

[0125] Among them, RVI S is the reference sample of the separation threshold, RVI summer is the RVI value of the summer image pixel, RVI winter is the RVI value of the winter image pixel.

[0126] In this step, in summer, the RVI value of deciduous forest is higher because the plants are growing more vigorously; while in winter, the RVI value of deciduous forest is lower because most of the deciduous plants have fallen leaves or are dormant. S It can stably characterize the differences between deciduous and evergreen forests. The RVI ratio of evergreen forests should be relatively stable and not significantly affected by seasonal changes, especially for coniferous forests. While the RVI ratio of shrubs may fluctuate, the RVI ratio of evergreen forests varies less between summer and winter. Broadleaf forests exhibit relatively greater variability, and the mean and standard deviation of the RVI ratio can be used to determine the threshold for separating evergreen and deciduous forests.

[0127] S1032. Obtain reference sample RVI of deciduous forest S The mean M RVI and standard deviation D RVI .

[0128] In this step, the reference sample RVI can be calculated using the following formula: S The mean M RVI and standard deviation D RVI :

[0129]

[0130] Among them, M RVI is the mean of the RVI ratios, D RVI is the standard deviation, and n is the number of samples.

[0131] S1033, based on the mean M RVI and standard deviation D RVI and multiple reference sample RVIs S The frequency histogram is drawn to obtain the reference sample RVI that obeys the normal distribution. S Frequency and RVI S Normal distribution curve.

[0132] In this step, the reference sample RVI is drawn based on the selected reference sample.S Assuming that these reference samples follow a uniform distribution, then according to the assumption of normal distribution, the distribution curve of the RVI ratio can be obtained. For details, please refer to Figure 2 .

[0133] S1034, according to RVI S Frequency and RVI S The normal distribution curve is used to set the discrimination rules, and the deciduous forest and evergreen forest are determined by the discrimination rules.

[0134] Figure 2 shows the RVI based on the reference sample S Schematic diagram of the discrimination rules between evergreen forest and deciduous forest based on frequency histogram. Figure 2 As shown, due to the selection of reference sample RVI S Following the principle of uniform distribution, the reference sample RVI is drawn S The frequency distribution of follows a normal distribution. In this embodiment, deciduous forest samples are selected as reference samples, where the discrimination rule is that if RVI S ∈[M RVI -D RVI ,M RVI +D RVI ], the forest feature of the pixel is identified as deciduous forest. S ∈[0,M RVI -2D RVI ], then the forest feature of the pixel is identified as evergreen forest. The interval for re-judgment is [M RVI -2D RVI , M RVI -D RVI ] and [M RVI +D RVI , M RVI +2D RVI ], the intervals to be re-determined are classified again using the mixed pixel decomposition method combined with texture features based on the pixels of the determined category as training samples. Texture features: In remote sensing image classification, texture features are used to capture spatial changes and patterns in images. This information helps to further distinguish different forest types (such as evergreen forests and deciduous forests). In this embodiment, in order to perform classification more accurately, the classified pixels are re-analyzed in combination with texture features and mixed pixel decomposition methods. Mixed pixel decomposition methods (such as the above-mentioned linear spectral mixture model) can process multiple components in mixed pixels, thereby improving the accuracy of classification.

[0135] S104. Savitzaky-Golay filter transformation and spectral differential transformation are performed on the forest dataset to obtain an NDVI transformed image set; the NDVI transformed image set indicates the time-series-based remote sensing phenological characteristics of plants, and the time-series-based remote sensing phenological characteristics indicate the growth and development status and cyclical changes of plants in different seasons.

[0136] It can be understood that time-series-based remote sensing phenological characteristics refer to time-series variation information obtained through remote sensing technology that reflects the growth cycle and ecological characteristics of plants. These characteristics mainly include changes in the reflectance spectrum during the growth, flowering, and withering of plants, and are usually used to analyze the phenological periods of plants (such as germination in spring, vigorous growth in summer, and leaf fall in autumn). Among them, plant phenology refers to the occurrence time and duration of each seasonal stage in the plant growth cycle (such as germination, flowering, and fruiting). Time-series remote sensing phenological characteristics help researchers monitor plant growth status, especially changes in vegetation in different seasons or climatic conditions, through remote sensing indices such as NDVI. For example, when vegetation recovers in spring, the NDVI value is usually higher; while in winter or drought periods, the NDVI value will decrease. By analyzing these time-series data, it is possible to infer the growth cycle of plants, suitable habitats, etc.

[0137] In some implementations, step S104 may be implemented by the following steps:

[0138] S1041. Based on the phenological information of dominant tree species in the study area, the Savitaky-Golay algorithm was used to perform polynomial filtering transformation on the NDVI time series data to obtain smoothed and denoised NDVI time series spectral curves of dominant tree species.

[0139] Among them, phenological information includes seasonal characteristics of dominant tree species, which indicates the canopy changes of each tree species during the period of leaf bud opening, leaf expansion, growth, full color leaf change and end of leaf fall. Figure 3 . Figure 3 Shows the seasonal characteristics of dominant tree species in the study area.

[0140] In this step, the Savitzky-Golay algorithm (also known as the SG filtering algorithm) performs a polynomial filtering transformation on the NDVI (Normalized Difference Vegetation Index) time series data to smooth and denoise the NDVI time series data, thereby enhancing the signal characteristics of the NDVI time series data and better classifying tree species.

[0141] It should be noted that the SG filtering algorithm is a filtering method based on local polynomial least squares in the time domain and is widely used for smoothing and denoising data streams. In the time series data analysis of this embodiment, the SG filtering algorithm is used to remove noise while maintaining the main characteristics of the signal. The key to the SG filtering algorithm is to weight the load points within the window, and the weights are obtained by least squares fitting of a given high-order polynomial. It can more effectively retain the characteristics of the signal while filtering and smoothing. After testing at actual sampling points, in this embodiment, the SG filtering algorithm with a smoothing window of 4 and a convolution dimension of 2 is finally adopted.

[0142] For example, the Savitaky-Golay algorithm is used to perform polynomial filtering transformation on NDVI time series data, including:

[0143] For NDVI time series data Y(t), (t=0,1…,Y N-1 ) where N is a positive integer, and for its adjacent 2k+1 data points (where k is a positive integer), fit an l-order polynomial in the local area:

[0144] σ(t)=a0+a1t+a2t 2 +…+a l t l

[0145] Where σ(t) is the fitted polynomial, which represents the fitted value at time point t (i.e. the result of polynomial fitting); a l (l=0, 1, ..., l) are the coefficients of the polynomial, which need to be obtained through fitting; l is the order of the polynomial (a positive integer), which represents the power of the highest-order term in the polynomial.

[0146] In this model, suppose we have a set of time series data Y(t), where t represents time (from 0 to N-1) and Y(t) is the observed value at time t. For each data point Y(t) in the time series, consider the adjacent data points around it. Specifically, 2k+1 data points are selected, where k is a positive integer indicating the number of adjacent data points to each data point. In other words, if we want to fit a data point Y(t), we will use the k points to its left and right (including itself, a total of 2k+1 points) to perform polynomial fitting.

[0147] Obtaining polynomial coefficient a based on local polynomial least squares fitting l , so that the square error between the fitting polynomial and the original data in the local area is minimized:

[0148]

[0149] Where E is the square error, Cj is the weighting coefficient, Y(t+j) is the actual observation value at time point t+j, and σ(t+j) is the actual fitting value at time point t+j calculated by polynomial fitting.

[0150] The goal of fitting is to find these polynomial coefficients a0, a1, ..., a l , so that the fitting polynomial in the local region can minimize the fitting error. A common approach is to find these coefficients using the least squares method. The error is defined as: for each data point Y(t+j) in the local region, the squared difference between the fitted curve σ(t+j) and the actual observed value Y(t+j), and the best fit is obtained by minimizing this squared error.

[0151] Figure 4 The NDVI time series spectrum curve of dominant tree species after SG filtering transformation is shown. Figure 4 As shown, through the above polynomial fitting process, the smoothed data point Y can be obtained (t) , these smoothed data points Y (t) That is, the data after filtering and denoising, connecting the above data point Y (t) The NDVI time series spectral curve of the dominant tree species after SG filtering transformation can be obtained.

[0152] S1042. Based on the first-order differential transformation and the second-order differential transformation, the NDVI time series spectral curve of the dominant tree species is obtained after extracting different spectral parameters.

[0153] In this step, the calculation formulas for the first-order differential transformation and the second-order differential transformation are as follows:

[0154]

[0155] In the formula, R(λ i ) is the spectral reflectance value at band i, R(λ i )′,R(λ i )″ are the first-order and second-order values ​​of the spectrum between bands i and i+1, and Δλ is the step size between adjacent bands.

[0156] It can be understood that spectral differential transformation mathematically simulates the reflectance spectrum and calculates differential values ​​of different orders to extract different spectral parameters. Spectral differential transformation can eliminate some atmospheric effects and the image of vegetation environments such as soil, thereby better reflecting the essential characteristics of plants. Specifically, the spectral curve after first-order differential can enhance the slope changes of the spectral curve, reflecting the rapid changes in the spectral curve between different bands and highlighting the phase shift differences of long-term spectral curves. In other words, the changing trend of the spectral curve is more obvious over time, which is helpful for analyzing the dynamic changes in the plant growth cycle. The second-order differential transformation can effectively reflect the inflection points and curvature characteristics of the spectral curve, enhancing the differences between individual spectral curves. Through the second-order differential transformation, detailed features of the spectral curve (such as sharp changes in reflectance) can be effectively explored, thereby enhancing the differences between different spectral curves and improving the accuracy of plant feature extraction. The second-order differential transformation can reveal the changing trends of the curve, such as the peaks and valleys of the spectral reflectance values, and is suitable for analyzing more complex vegetation changes. The third-order differential transformation can theoretically further explore detailed characteristics, such as more subtle reflectance changes. It can capture more subtle changes and enhance the characteristic details of the spectrum. However, the calculation result of the third-order differential is related to the curvature of the original space vector (rotation of the curve), which may cause some unstable changes. Therefore, in this embodiment, the third-order differential transformation is not discussed in detail.

[0157] Figure 5 The NDVI time series spectrum curve of dominant tree species after SG filtering and differential transformation is shown. Figure 5 As shown in the figure, through spectral differential transformation, the spectral interference caused by factors such as atmosphere and soil can be eliminated and the reflection characteristics of the plant itself can be enhanced.

[0158] S105. Classify the tree species in the study area using a random forest model after combining the vegetation spectral index and the NDVI transformed image set as classification features.

[0159] In this step, the vegetation spectral index and NDVI transformed image sets are used as input variables for the random forest model to output tree species identification results, including identification of individual tree species in deciduous forests and evergreen forests. The random forest algorithm (RF) integrates multiple trees using the bagging concept of ensemble learning. The basic unit is the decision tree. A decision tree is a tree-like structure that divides data into different categories through nodes and branches. Each node represents a conditional judgment on a feature, branches represent different decision outcomes, and finally, leaf nodes represent the classification results.

[0160] For example, the random forest model uses a bootstrap method to randomly select N samples from the original training set as training samples. In this implementation, a 7:3 ratio is used for training sample selection. The remaining 30% of samples are called out-of-bag (OOB) data and can be used to estimate the performance of the random forest model. During the training process of each tree, due to the random selection nature of the bootstrap method, each training sample may be selected multiple times, and some samples may not be selected. These unselected samples are the out-of-bag data.

[0161] Optionally, the training samples include multi-season phenological characteristics, Sentinel-1 radar data, and red-edge regional bands. Multi-season phenological characteristics: Extract multi-season vegetation growth characteristics from time series images, such as NDVI values ​​and curve morphology in different seasons. Sentinel-1: Combined with Sentinel-1 SAR data (such as synthetic aperture radar imagery), the scattering characteristics of radar signals are used to improve classification accuracy. Red-edge regional bands: Combined with Sentinel-2 red-edge bands, the red-edge bands are used to perform sensitivity analysis on vegetation and distinguish different tree species. These features are input into the random forest model for training, and the classification accuracy can be effectively improved through multi-decision tree ensemble classification.

[0162] In some embodiments, the tree species identification method based on dense time-series images provided in the embodiments of the present application further includes S106 , performing accuracy verification on the tree species identification result.

[0163] Figure 6 The radar charts of the recognition accuracy of each dominant tree species based on different feature combinations are shown (SI: spectral index; SG: SG filter transform; FD: first-order differential transform; SD: second-order differential transform). Figure 6 As shown in the figure, accuracy verification can evaluate the tree species identification results from the perspectives of overall accuracy (OA) and kappa coefficient, producer's accuracy (PA), and user's accuracy (UA). The calculation formulas for each parameter are as follows:

[0164]

[0165] Where h i is the number of correctly classified samples of the i-th category; X i is the total number of classified pixels of the i-th category; T i is the total number of true pixels of the i-th category; N is the total number of classified pixels.

[0166] The calculated result of kappa is -1 to 1, but usually kappa falls between 0 and 1. It can be divided into five groups to represent different levels of consistency: 0.0 to 0.20 is very low consistency (slight), 0.20 to 0.40 is general consistency (fair), 0.40 to 0.60 is moderate consistency (moderate), 0.60 to 0.80 is high consistency (substantial) and 0.80 to 1 is almost perfect.

[0167] The tree species identification method based on dense time-series images provided in the embodiments of the present application involves a tree species identification method, apparatus, and device based on dense time-series images, and its application scenarios include but are not limited to:

[0168] (1) Dynamic update of forest resources. Tree species identification is an important part of forestry resource surveys, and accurate identification of tree species is very important for natural resource management. Conventional tree species identification mainly relies on field surveys, which is labor-intensive, time-consuming, and resource-intensive. This study uses remote sensing technology to make the acquisition of forest resources faster, more accurate, and more efficient. It plays a very important role in the dynamic monitoring of forestry resources and the classification and identification of tree species.

[0169] (2) Accurately estimate carbon storage. Accurately estimating tree and forest biomass is essential for assessing and tracking the current status of forest carbon pools and their rate of change, predicting the role and feedback of forests in regional and global carbon cycles, and providing an important basis for formulating policies to address climate change. Tree species diversity has different effects on carbon storage for different types of tree species. The calculation of carbon storage requires the establishment of a set of equations based on different tree species. Therefore, clarifying the distribution of regional tree species is the basis for accurately estimating carbon storage.

[0170] (3) Real-time monitoring of precious tree species. The tree species mapping is used to screen precious tree species, clarify and update the spatial location of precious tree species in real time, provide real-time information for forest protectors, and facilitate the timely and effective formulation of protection measures.

[0171] (4) Monitoring of forest tree species diversity. Remote sensing technology is gradually becoming an emerging means of rapid large-scale monitoring of forest biodiversity, providing a strong guarantee for the rapid extraction of spatial pattern information of tree species diversity. Large-scale tree species mapping provides a data basis for regional forest species diversity detection.

[0172] Example 1

[0173] Figure 7 This is a flow chart of the application scenario provided in Example 1 of this application. Figure 7As shown in the figure, five typical tree species in a region (Pinus tabulaeformis, Quercus liaotungensis, Populus davidiana, Larix principis-rupprechtii, and others) were selected as the subjects. The images used for tree species extraction included 66 Sentinel-2 remote sensing images from 2019 to 2022. The image preprocessing and tree species identification methods used included the Linear Spectral Mixture Model (LSMM), the RVI threshold method, the Savitzaky-Golay filter transform, the spectral differential transform, and the random forest algorithm. The specific steps are as follows:

[0174] S201. Obtain original data.

[0175] For example, the original data includes Sentinel-2, terrain data, sample data and Sentinel-1. Among them, the original remote sensing image data of Sentinel-1 and Sentinel-2 are all from the European Space Agency. A total of 66 scenes of Sentinel-2 remote sensing image data of the Huodong mining area in the years participating in the evaluation were obtained through the GEE platform. The terrain data uses the NASADEM dataset. NASADEM is a reprocessing of digital elevation model data generated by the Shuttle Radar Topography Mission (SRTM) data. By incorporating data mainly from the Ice, Cloud and Land Elevation Satellite (ICESat), the Earth Science Laser Altimetry System (GLAS) and the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) instruments, NASADEM can provide more accurate global terrain data. The sample data comes from existing forest inventory data and field survey data, of which 70% are used as training samples for training classification models, and 30% are used as validation samples for random forest model accuracy verification. Field survey data were used as field data. A systematic sampling strategy of random points was adopted to design 30m*30m square plots. The diameter and height of trees with a breast diameter greater than 4m were measured and the tree species was recorded. At the same time, handheld laser radar (SLAM) was used for scanning. A total of 412 sample plots were collected.

[0176] S202: Preprocessing.

[0177] First, we preprocessed the Sentinel-2 time series imagery. Since the pixel resolution of Sentinel-2 bands B5, B6, B7, B8A, B11, and B12 is 20 meters, we resampled them to 10 meters using bicubic interpolation. Bicubic interpolation offers edge enhancement and can effectively preserve image fine structure, but it is computationally intensive. Using the GEE platform effectively addressed this weakness. Next, we reprojected bands B2, B3, B4, and B8, along with their resampled bands, to the EPSG:4326 coordinate system and fused the bands into a new image.

[0178] To fully utilize the high spectral resolution of Sentinel-2, vegetation indices (Table 1) were constructed using information from different bands of the image. These indices were then combined with the 10 preprocessed bands of Sentinel-2 to construct 12 characteristic variables as shown in Table 1. These characteristic variables help improve the accuracy of tree species classification.

[0179] For example, the raw remote sensing image data can also be pre-processed by the satellite platform, including cloud removal, terrain correction, and image cropping. Optionally, the cloud removal can use the Simple Cloud Score algorithm; the terrain correction can use the SCS+C correction method based on the DEM Lambertian reflectance model; and the image cropping can use the Quality Mosaic algorithm.

[0180] S203. Use the linear spectral mixture model combined with terrain data to classify the land use of the study area into forest land, buildings, water bodies and other land features.

[0181] S204: Based on the extracted forest land, the regional forest land is pre-classified using the RVI threshold method. 280 pure deciduous forest sample points are found and RVI is drawn. S Frequency histogram and calculation of RVI S The mean and standard deviation of Figure 2 Classification of evergreen forests and deciduous forests is carried out based on the discrimination rules.

[0182] S205, performing Savitzaky-Golay filter transformation and spectral differential transformation.

[0183] All cloud-removed images from 2019 to 2022 were used. The NDVI time series spectral curves of the canopy of the five dominant tree species after SG algorithm filtering and spectral differential transformation were calculated as follows: Figure 4 and Figure 5 .

[0184] S206. Tree species classification and accuracy evaluation.

[0185] Based on the result obtained in step S205 Figure 4 and Figure 5 The NDVI time series spectral curve information of the canopy of dominant tree species is used as the input variable of the random forest model to identify the various tree species in deciduous forests and evergreen forests.

[0186] The accuracy of tree species classification results was verified. The tree species identification results were evaluated based on overall accuracy, kappa coefficient, producer accuracy, and user accuracy.

[0187] In this embodiment, the overall accuracy and Kappa coefficient of tree species classification based on the dominant tree species feature combinations of different time series transformation data are shown in Table 2:

[0188] Table 2 Overall accuracy and Kappa coefficient of dominant tree species feature combinations based on different time series transformation data

[0189]

[0190]

[0191] The Kappa coefficient of the tree species extraction accuracy in Table 2 shows that the method of the present application has a high degree of consistency (substantial) in the tree species classification results.

[0192] It is understandable that the size of the serial number of each step in the above-mentioned embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application. In addition, in some possible implementations, the steps in the above-mentioned embodiments can be selectively executed according to actual conditions, and can be partially executed or fully executed, which is not limited here. In addition, all or part of any features of any of the above-mentioned embodiments can be freely and arbitrarily combined without contradiction; the combined technical solutions are also within the scope of this application.

[0193] like Figure 8 As shown, the embodiment of the present application also provides a tree species identification device 300 based on dense time-series images, including: an acquisition module 301 and a processing module 302. The acquisition module 301 is used to obtain dense long-time series multispectral remote sensing images and terrain data of the study area; the dense long-time series multispectral remote sensing images include all available remote sensing forest images with less than 20% cloud cover, such as 2019-2022; the processing module 302 is used to perform image band fusion based on the dense long-time series multispectral remote sensing images to construct a vegetation spectral index; the vegetation spectral index is used to indicate the dynamic change characteristic variables of vegetation growth within the year; the linear spectral mixture model is used in combination with terrain data to classify the land use of the objects in the dense long-time series multispectral remote sensing images of the study area into buildings, water bodies and other objects, and woodlands. A dataset was obtained, and pre-classification was performed based on the extracted woodland dataset. The woodland dataset was divided into evergreen forest and deciduous forest using the RVI threshold method combined with the linear spectral mixture model; Savitzaky-Golay filter transformation and spectral differential transformation were performed on the woodland dataset to obtain an NDVI transformed image set; the NDVI transformed image set indicated the time-series-based remote sensing phenological characteristics of plants, and the time-series-based remote sensing phenological characteristics indicated the growth and development status and cyclical changes of plants in different seasons; the tree species in the study area were classified using a random forest model after combining the vegetation spectral index and the NDVI transformed image set as classification features.

[0194] It should be understood that both the acquisition module 301 and the processing module 302 can be implemented by software or hardware. For example, the implementation of the acquisition module 301 will be described below using the acquisition module 301 as an example. Similarly, the implementation of the processing module 302 can refer to the implementation of the acquisition module 301.

[0195] As an example of a software functional unit, the acquisition module 301 may include code running on a computing instance. The computing instance may include at least one of a physical host (computing device), a virtual machine, and a container. Furthermore, the computing instance may be one or more. For example, the acquisition module 301 may include code running on multiple hosts / virtual machines / containers. It should be noted that the multiple hosts / virtual machines / containers used to run the code may be distributed in the same region or in different regions. Furthermore, the multiple hosts / virtual machines / containers used to run the code may be distributed in the same availability zone (AZ) or in different AZs, each AZ including one data center or multiple geographically close data centers. Typically, a region may include multiple AZs.

[0196] Similarly, multiple hosts / virtual machines / containers used to run the code can be distributed within the same virtual private cloud (VPC) or across multiple VPCs. Typically, a VPC is set up within a region. Cross-region communication between two VPCs within the same region, or between VPCs in different regions, requires a communication gateway within each VPC to interconnect the VPCs.

[0197] As an example of a hardware functional unit, acquisition module 301 may include at least one computing device, such as a server. Alternatively, acquisition module 301 may be implemented using an application-specific integrated circuit (ASIC) or a programmable logic device (PLD). The PLD may be implemented using a CPLD, FPGA, GAL, or any combination thereof.

[0198] The multiple computing devices included in acquisition module 301 can be distributed in the same region or in different regions. The multiple computing devices included in acquisition module 301 can be distributed in the same AZ or in different AZs. Similarly, the multiple computing devices included in acquisition module 301 can be distributed in the same VPC or in multiple VPCs. The multiple computing devices can be any combination of servers, ASICs, PLDs, CPLDs, FPGAs, GALs, and other computing devices.

[0199] It should be understood that the above-mentioned device is used to execute the method in the above-mentioned embodiment. The implementation principle and technical effect of the corresponding program module in the device are similar to those described in the above-mentioned method. The working process of the device can refer to the corresponding process in the above-mentioned method and will not be repeated here.

[0200] Figure 9 A computing device provided in an embodiment of the present application, such as Figure 9 As shown, the computing device 70 includes: at least one memory 702 for storing programs; at least one processor 701 for executing the programs stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method for tree species identification based on remote sensing images as described in any of the above embodiments.

[0201] An embodiment of the present application provides a computer storage medium 703, in which instructions are stored. When the instructions are executed on a computer, the computer executes the method for tree species identification based on remote sensing images as described in any of the above embodiments.

[0202] It is understood that the processor in the embodiments of the present application may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0203] The present application also provides a computer-readable storage medium. The computer-readable storage medium is used to store computer program instructions. When the computer program instructions are executed on a computing device, the computing device executes the data processing method involved in the above embodiment. The computer-readable storage medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a magnetic tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive).

[0204] The present application also provides a computer program product comprising instructions. The computer program product may be software or a program product comprising instructions that can be run on a computing device or stored in any available medium. When the computer program product is run on at least one computing device, the at least one computing device executes the data processing method.

[0205] Finally, it should be noted that the above embodiments are merely illustrative of the technical solutions of the present application. Those skilled in the art should understand that, although the present application has been described in detail with reference to the aforementioned embodiments, the technical solutions described in the aforementioned embodiments may be modified or some of the technical features thereof may be replaced with equivalents. However, such modifications or replacements do not deviate from the spirit and scope of the technical solutions in the various embodiments of the present application.

Claims

1. A tree species identification method based on dense time series images, characterized in that: include: Acquire dense long-term multispectral remote sensing images and terrain data of the study area; the dense long-term multispectral remote sensing images include all available remote sensing images with cloud cover less than 20%; Performing image band fusion based on the dense long-term multispectral remote sensing image to construct a vegetation spectral index; the vegetation spectral index is used to indicate the characteristic variables of the dynamic changes in vegetation growth within the year; The linear spectral mixture model was used in combination with terrain data to classify the land use of objects in the dense long-term multispectral remote sensing images of the study area into buildings, water bodies and other objects, and forest datasets. The extracted forest dataset was pre-classified using the RVI threshold method combined with the linear spectral mixture model to classify the forest dataset into evergreen forest and deciduous forest. Savitzaky-Golay filter transformation and spectral differential transformation are performed on the forest dataset to obtain an NDVI transformed image set; the NDVI transformed image set indicates the time-series-based remote sensing phenological characteristics of plants, and the time-series-based remote sensing phenological characteristics indicate the growth and development status and periodic changes of plants in different seasons; The tree species in the study area are classified using a random forest model after combining the vegetation spectral index and NDVI transformed image set as classification features.

2. The method according to claim 1, characterized in that The step of performing image band fusion based on the dense long-time series multispectral remote sensing image to construct a vegetation spectral index includes: The dense long-time multispectral remote sensing images and terrain data are preprocessed to obtain forest growth images, and the preprocessing includes image registration, resampling, time series and cloud screening, radiation correction, geometric correction, band synthesis, cloud and shadow masking, and image cropping.

3. The method according to claim 2, characterized in that The resampling comprises: The terrain data were resampled to a spatial resolution of 10 m using a linear sampling method; The dense long-term multispectral remote sensing image is resampled to 10m by using bicubic interpolation method with a pixel resolution of 20m.

4. The method according to any one of claims 1 to 3, characterized in that The land use classification of objects in dense long-term multispectral remote sensing images of the study area using a linear spectral mixture model combined with terrain data includes: The linear spectral mixture model refers to the spectral reflectance of a pixel in a certain band. It is a linear combination of the reflectance of each component constituting the pixel with its area ratio as the weight coefficient, which is expressed as: Where: R iλ is the spectral reflectance of the i-th pixel in the λ band; r kλ is the spectral reflectance of the kth basic component in the uth band; f ki is the abundance of the kth endpoint element in the i-th pixel; ξ iλ It is the residual value, representing the multiple reflections and transmissions of light between pixel components, which has a nonlinear mixing effect; n is the number of terminal elements, and m represents the number of basic components that make up the pixel.

5. The method according to claim 4, characterized in that The vegetation spectral index includes a ratio vegetation index (RVI); and the pre-classification based on the extracted forest data set, which divides the forest data set into evergreen forest and deciduous forest, includes: The RVI values ​​of sample points in the deciduous forest in winter and summer from the field survey data were randomly selected for ratio processing, and the minimum value of the relatively stable RVI ratio of the two images was obtained as the separation threshold: Among them, RVI S is the reference sample of the separation threshold, RVI summer is the RVI value of the summer image pixel, RVI winter is the RVI value of the winter image pixel; Obtaining reference sample RVI for deciduous forests S The mean M RVI and standard deviation D RVI ; Based on the mean M RVI and standard deviation D RVI and multiple reference sample RVIs S The frequency histogram is drawn to obtain the reference sample RVI that obeys the normal distribution. S Frequency and RVI S Normal distribution curve; According to RVI S Frequency and RVI S The normal distribution curve sets a discrimination rule, and deciduous forests and evergreen forests are determined by the discrimination rule.

6. The method according to claim 4, characterized in that The vegetation spectral index includes time series data of the Normalized Difference Vegetation Index (NDVI); the filtering transformation and spectral differential transformation are performed on the forest data set to obtain an NDVI transformation image set, including: Based on the phenological information of dominant tree species in the study area, the Savitaky-Golay algorithm was used to perform polynomial filtering on the NDVI time series data to obtain smoothed and denoised NDVI time series spectral curves of dominant tree species. The phenological information includes seasonal characteristics of dominant tree species, which indicates canopy changes in the leaf bud opening period, leaf expansion period, growth period, leaf full color period, and late leaf fall period of each tree species. Based on the first-order differential transformation and the second-order differential transformation, the NDVI time series spectral curves of dominant tree species were obtained after extracting different spectral parameters.

7. The method according to claim 6, characterized in that The polynomial filtering transformation of NDVI time series data using the Savitaky-Golay algorithm includes: For NDVI time series data Y(t), (t=0,1…,Y N-1 ) where N is a positive integer, and for its adjacent 2k+1 data points (where k is a positive integer), fit an l-order polynomial in the local area: σ(t)=a0+a1t+a2t 2 +t…+a l t l Where σ(t) is the fitting polynomial; a l (l=0, 1, ..., l) are coefficients of the polynomial; Obtaining polynomial coefficient a based on local polynomial least squares fitting l , so that the square error between the fitting polynomial and the original data in the local area is minimized: Where E is the square error, C j is the weighting coefficient, Y(t+j) is the actual observation value at time point t+j, and σ(t+j) is the actual fitting value at time point t+j calculated by polynomial fitting; Through the above polynomial fitting process, the smoothed data point Y can be obtained (t) , these smoothed data points Y (t) That is, the data after filtering and denoising is used to obtain the NDVI time series spectral curve of the dominant tree species after SG filtering transformation.

8. The method according to claim 1, characterized in that The NDVI time series spectral curve of the dominant tree species after extracting different spectral parameters based on the first-order differential transformation and the second-order differential transformation includes: The calculation formulas for first-order differential transform and second-order differential transform are as follows: In the formula, R(λ i ) is the spectral reflectance value at band i, R(λ i )′,R(λ i )″ are the first-order and second-order values ​​of the spectrum between bands i and i+1, and Δλ is the step size between adjacent bands.

9. The method according to any one of claims 1 to 3, characterized in that The method further includes: performing accuracy verification on the tree species identification result, including: The tree species identification results were evaluated from the perspectives of overall accuracy and kappa coefficient; The overall accuracy evaluation is to compare the pixel-level tree species recognition results of long-term remote sensing image data with the ground truth to evaluate the accuracy of the classification results. The calculation formula for the overall accuracy is as follows: Overall accuracy = (TP+TN) / (TP+TN+FP+FN) TP represents the number of samples that are truly positive and correctly predicted as positive, TN represents the number of samples that are truly negative and correctly predicted as negative, FP represents the number of samples that are truly negative but incorrectly predicted as positive, and FN represents the number of samples that are truly positive but incorrectly predicted as negative. The Kappa coefficient is a statistical indicator for evaluating the consistency between classifiers or evaluators. The calculation formula of Kappa coefficient is as follows: Among them, Po is the sum of the number of correctly classified samples in each category divided by the total number of samples, that is, the overall classification accuracy; assuming that the number of true samples in each category is a1, a2, ..., a C , and the predicted number of samples of each class are b1, b2, ..., b C The total number of samples is n, then: The calculated result of kappa is -1 to 1, but usually kappa falls between 0 and 1. It can be divided into five groups to represent different levels of consistency: 0.0 to 0.20 is very low consistency (slight), 0.20 to 0.40 is general consistency (fair), 0.40 to 0.60 is moderate consistency (moderate), 0.60 to 0.80 is high consistency (substantial) and 0.80 to 1 is almost perfect.

10. A tree species identification device based on dense time-series images, characterized in that: include: An acquisition module, the acquisition module is used to acquire dense long-term multispectral remote sensing images and terrain data of the study area; the dense long-term multispectral remote sensing images include all available remote sensing forest images with cloud cover less than 20%; A processing module is used to perform image band fusion based on the dense long-term multispectral remote sensing image to construct a vegetation spectral index; the vegetation spectral index is used to indicate the characteristic variables of the dynamic changes in vegetation growth within the year; The linear spectral mixture model combined with terrain data was used to classify the land use of the features in the dense long-term multispectral remote sensing images of the study area into buildings, water bodies and other features, as well as forest datasets. The extracted forest dataset was pre-classified, and the RVI threshold method combined with the linear spectral mixture model was used to divide the forest dataset into evergreen forest and deciduous forest. Savitzaky-Golay filtering transformation and spectral differential transformation were performed on the forest dataset to obtain the NDVI transformed image set. The NDVI transformed image set indicated the time-series-based remote sensing phenological characteristics of plants, and the time-series-based remote sensing phenological characteristics indicated the growth and development status and periodic changes of plants in different seasons. The vegetation spectral index and the NDVI transformed image set were combined as classification features and then the tree species in the study area were classified using a random forest model.

Images