A method and system for tea garden identification based on multi-source remote sensing data

Abstract

This invention discloses a tea garden identification method and system based on multi-source remote sensing data. The method includes the following steps: S1: acquiring temporal radar data and optical remote sensing data of the target area; S2: extracting key growth periods of the tea garden based on radar data, including the beginning and end of the growing season; S3: calculating vegetation index VI and water index WI using optical data, and selecting the VI and WI combination with optimal discrimination through a time-weighted dynamic time warping method; S4: determining the tea garden identification index as follows: where Ω1 is a linear operation, Ω2 is a nonlinear operation, T represents the tea garden sample, G represents the evergreen vegetation sample, and DTW is the dynamic time warping method; S5: realizing tea garden identification based on the tea garden identification index and an ensemble learning model.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing image processing technology, specifically to a tea garden identification method and system based on multi-source remote sensing data. Background Technology

[0002] Tea, as an important economic crop and one of the world's three major beverages, plays a significant role in promoting regional economic development, especially in developing countries. Accurate and efficient identification and monitoring of the spatial distribution of tea gardens is crucial for scientifically guiding land use planning, rational tea planting, disease control, yield estimation, and ecological environmental protection.

[0003] Traditional methods of obtaining tea garden area data rely on field surveys and agricultural censuses, which suffer from significant drawbacks such as being time-consuming and labor-intensive, highly subjective, having delayed updates, and lacking detailed spatial information. In contrast, satellite remote sensing technology, with its advantages of wide monitoring range, strong spatiotemporal continuity, and relatively low cost, has become the main means of land cover monitoring and has been widely used in crop classification and identification. Existing research attempts to use remote sensing technology to identify tea gardens, but faces two major technical challenges:

[0004] Insufficient temporal completeness of optical data: Especially in major tea-producing areas such as southern China, frequent cloudy and rainy weather causes optical satellite imagery to be frequently obscured by clouds, making it difficult to obtain complete and high-quality time-series data. Although time-series interpolation methods can partially fill in the data gaps, they are unable to capture the unique, short-cycle (e.g., 5-8 day picking period) phenological characteristics of tea gardens. Research on fusing radar (SAR) data with optical data to overcome cloud obscuration has made some progress, but its models are usually based on the assumption of single-phase multimodal translation, ignoring the seasonal and temporal evolution of ground object spectra, resulting in a lack of consistency and stability in the reconstruction results over time. In addition, existing methods often attempt to simulate up to 13 optical bands with limited radar channels (e.g., VV / VH), resulting in a severe lack of information and difficulty in accurately recovering the spectral details and relative spectral shapes of each band.

[0005] Tea gardens suffer from severe spectral confusion with evergreen vegetation: Tea gardens exhibit highly similar spectral reflectance characteristics to other evergreen vegetation (such as forests and shrubs) in the visible to near-infrared bands, stemming from similar chlorophyll content, water status, and canopy structure. This similarity is pervasive across different climates, terrains, and management models, and remains stable across multiple temporal observations, leading to the loss of key discriminative features (such as red edges and near-infrared regions). This severely weakens the ability of classification algorithms to extract effective features and their generalization ability in complex environments, becoming a critical bottleneck restricting the accurate identification of large-scale tea gardens. While introducing hyperspectral data can enhance the capture of subtle physiological and biochemical differences, its high acquisition cost, limited coverage, high data redundancy, and complex processing make it difficult to meet the practical needs of large-scale, long-term tea garden monitoring.

[0006] Index-based methods have gained attention in crop identification due to their ability to enhance the spectral differences between target crops and other land cover, and have shown good results and cost-effectiveness in mapping other crops (such as soybeans, rapeseed, potatoes, and winter wheat). However, to date, there is no effective identification index designed for the unique and complex spectral-temporal characteristics of tea gardens that can reliably distinguish tea gardens from other spectrally similar evergreen vegetation on a large scale.

[0007] In general, the main challenges facing existing remote sensing technologies in the field of tea garden identification are: how to overcome cloud interference to obtain complete, high-quality time-series data that fully reflects the key phenological periods (especially short-period characteristics) of tea gardens; and how to extract features with high discriminative power, strong robustness, and good generalization ability from complex and variable spectral information to effectively solve the serious spectral confusion problem between tea gardens and evergreen vegetation. Developing new technologies that can solve these two major problems simultaneously is of urgent need and great significance for achieving efficient, accurate, and large-scale tea garden identification and dynamic monitoring. Summary of the Invention

[0008] To address the technical challenges in existing technologies, such as insufficient temporal integrity of optical data, severe spectral confusion between tea gardens and evergreen vegetation, and the lack of effective identification indices designed for the unique and complex spectral-temporal characteristics of tea gardens, this invention proposes a tea garden identification method and system based on multi-source remote sensing data. By establishing a "radar timing-optical classification" collaborative framework and phenological driving indices, the stability of large-scale tea garden identification is significantly improved.

[0009] To achieve this goal, the present invention adopts the following technical solution.

[0010] A method for identifying tea gardens based on multi-source remote sensing data, comprising the following steps:

[0011] S1: Acquire temporal radar data and optical remote sensing data of the target area;

[0012] S2: Extract key growth periods of tea gardens based on radar data, including the beginning and end of the growing season;

[0013] S3: Calculate vegetation index (VI) and water index (WI) using optical data, and select the VI and WI combination with the best discrimination by time-weighted dynamic time warping method.

[0014] S4: Determine the tea garden identification index as follows:

[0015] Where Ω1 is a linear operation, Ω2 is a nonlinear operation, T represents the tea garden sample, G represents the evergreen vegetation sample, and DTW is the dynamic time warping method.

[0016] S5: Tea garden identification is achieved based on the tea garden identification index and ensemble learning model.

[0017] Furthermore, in the tea garden identification method based on multi-source remote sensing data of the present invention, the extraction of key growth stages of tea gardens based on radar data includes:

[0018] a) Perform radiometric calibration, geometric correction and filtering on radar data to construct a time series of VV / VH polarization backscattering coefficients;

[0019] b) For the VV / VH polarization backscattering coefficient time series, filter smoothing is performed. The formula for calculating the filter smoothing is:

[0020] ,

[0021] in, These are the values ​​after filtering and smoothing. For the original value, The convolution coefficients are... The normalization coefficient is... is the half-width of the filtering window, and i is the offset relative to the center j of the sliding window, with a value ranging from -m to m;

[0022] The start-of-season SOS and end-of-season EOS are determined based on the extreme points of the first derivative, as follows:

[0023] ,

[0024] ,

[0025] The argmax function is used to find the maximum value of a function within a given range. This is the change in the backscattering coefficient. This represents the corresponding time interval.

[0026] Furthermore, in the tea garden identification method based on multi-source remote sensing data of the present invention, the selection of the VI and WI combination with the best discrimination through the time-weighted dynamic time warping method includes:

[0027] Constructing the Tea Garden Timeline With evergreen vegetation time sequence The weighted distance matrix is:

[0028] ,in Point and Weighted distance between them;

[0029] Calculate the elements of the cumulative distance matrix:

[0030] ,

[0031] Construct the cumulative distance matrix ,in This represents the distance from the start of the sequence to the current position. The minimum cumulative distance;

[0032] The output time-weighted dynamic time warp distance is: .

[0033] In addition, in the tea garden identification method based on multi-source remote sensing data of the present invention, the vegetation index includes at least one of the normalized vegetation index NDVI, enhanced vegetation index EVI, and ratio vegetation index RVI, and the water index includes at least one of the terrestrial water index LSWI, normalized water index NDWI, and modified normalized difference water index mNDWI.

[0034] In addition, the tea garden identification method based on multi-source remote sensing data of the present invention also includes a pre-extraction step of evergreen vegetation:

[0035] a) For each of the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Ratio Vegetation Index (RVI), Land Water Index (LSWI), Normalized Water Index (NDWI), and Modified Normalized Difference Water Index (mNDWI), four time-series statistical features are extracted: annual mean, maximum value, minimum value, and annual variation, and a 24-dimensional multi-temporal feature space is constructed.

[0036] b) Combine random forest, support vector machine and extreme gradient boosting methods to classify land cover in the target area and extract evergreen vegetation areas;

[0037] c) Morphological processing methods, including opening and closing operations and region connectivity analysis, are used to remove isolated misclassified pixels and enhance the spatial continuity of the results.

[0038] Furthermore, in the tea garden identification method based on multi-source remote sensing data of the present invention, the tea garden identification is achieved using an ensemble learning model, including:

[0039] Integrating the probability outputs of three classifiers: Random Forest, Support Vector Machine, and Extreme Gradient Boosting:

[0040] ,

[0041] in, It is the final classification probability. is the probability prediction of the i-th classifier, and M is the number of classifiers;

[0042] A feature space is constructed by combining topographic parameters such as elevation, slope, and aspect.

[0043] In addition, in the tea garden identification method based on multi-source remote sensing data of the present invention, the radar data is Sentinel-1 C-band synthetic aperture radar SAR data, and the optical data is Sentinel-2 multispectral data.

[0044] In addition, the present invention also includes a tea garden identification system based on multi-source remote sensing data, which includes a data acquisition module, a phenological analysis module, an index construction module, a classification module, and an output module, wherein...

[0045] The data acquisition module is used to acquire time-series radar data and optical remote sensing data of the target area;

[0046] The phenology analysis module is used to extract key growth stages in tea gardens based on radar data;

[0047] The index construction module is used to calculate the vegetation index (VI) and water index (WI) using optical data, and to select the optimal combination of VI and WI using a time-weighted dynamic time warping method; and to determine the tea garden identification index as follows:

[0048] Where Ω1 is a linear operation, Ω2 is a nonlinear operation, T represents the tea garden sample, G represents the evergreen vegetation sample, and DTW is the dynamic time warping method.

[0049] The classification module identifies tea gardens based on a tea garden identification index and an ensemble learning model.

[0050] The output module is used to generate a spatial distribution map of tea gardens based on the tea garden identification results.

[0051] Furthermore, in the tea garden identification system based on multi-source remote sensing data of the present invention, the phenological analysis module extracts the key growth stages of the tea garden based on radar data, including:

[0052] a) Perform radiometric calibration, geometric correction and filtering on radar data to construct a time series of VV / VH polarization backscattering coefficients;

[0053] b) For the VV / VH polarization backscattering coefficient time series, filter smoothing is performed. The formula for calculating the filter smoothing is:

[0054] ,

[0055] in, These are the values ​​after filtering and smoothing. For the original value, The convolution coefficients are... The normalization coefficient is... is the half-width of the filtering window, and i is the offset relative to the center j of the sliding window, with a value ranging from -m to m;

[0056] The start-of-season SOS and end-of-season EOS are determined based on the extreme points of the first derivative, as follows:

[0057] ,

[0058] ,

[0059] The argmax function is used to find the maximum value of a function within a given range. This is the change in the backscattering coefficient. This represents the corresponding time interval.

[0060] Furthermore, in the tea garden identification system based on multi-source remote sensing data of the present invention, the index construction module is used to calculate the vegetation index VI and the water body index WI using optical data, and to screen the VI and WI combination with the best discrimination through a time-weighted dynamic time warping method, including:

[0061] Constructing the Tea Garden Timeline With evergreen vegetation time sequence The weighted distance matrix is:

[0062] ,in Point and Weighted distance between them;

[0063] Calculate the elements of the cumulative distance matrix:

[0064] ,

[0065] Construct the cumulative distance matrix ,in This represents the distance from the start of the sequence to the current position. The minimum cumulative distance;

[0066] The output time-weighted dynamic time warp distance is: .

[0067] The technical effects of this invention include the following.

[0068] The Time-Weighted Dynamic Time Warping (TWDTW) method not only considers the similarity of time series shapes but also incorporates time dimension matching constraints into the calculation, making the technical solution of this invention more sensitive to seasonal changes. This characteristic is particularly important when distinguishing vegetation types with similar spectral characteristics but different phenological rhythms (such as tea gardens and other evergreen vegetation). TWDTW more accurately quantifies the distance between time series by finding the optimal alignment path between two time series while penalizing excessive distortion on the time axis. First, time series of multiple vegetation indices and water indices need to be calculated. Then, the index time series of tea garden samples and other evergreen vegetation are extracted separately to ensure the integrity and comparability of the time series data. The time series distance between tea garden and other evergreen vegetation samples is calculated using TWDTW, and the vegetation index (VI) and water index (WI) with the largest TWDTW distance are selected through statistical analysis. These indices will be used for subsequent TPRI index construction.

[0069] The rationality of the TWDTW distance maximization principle is reflected on two levels: From a phenological perspective, the TWDTW distance can accurately measure the temporal differences between tea gardens and other evergreen vegetation throughout their growth cycle. A larger TWDTW distance means that the constructed tea garden identification index can better capture the unique growth patterns of tea gardens. From the perspective of index construction objectives, since the standard for selecting the optimal VI and WI is the maximum TWDTW distance, their combination (i.e., TPRI) should produce an even larger TWDTW distance. Otherwise, it indicates that this combination weakens rather than enhances the distinguishing ability of the original index, violating the original intention of index construction. Therefore, by verifying whether the TWDTW distance of the tea garden identification index exceeds that of all individual indices, the consistency of the evaluation criteria is ensured, and the combination process is guaranteed to truly play an optimization role.

[0070] This invention systematically explores and evaluates the discriminative power of different index combinations, ultimately selecting the index combination that best distinguishes tea gardens from evergreen forests, and thereby constructing the TPRI index specifically for identifying tea gardens. The choice to select one index from VI and one from WI, rather than simply choosing the two with the largest D_index, is based on the complementarity of these two types of indices in reflecting land cover characteristics: VI mainly reflects physiological characteristics such as vegetation biomass and photosynthesis, while WI focuses on reflecting the water content of vegetation and soil.

[0071] The Tea Garden Identification Index (TPRI) is constructed based on the following core ideas: First, it utilizes the complementarity of the selected VI and WI indices to comprehensively consider vegetation growth status and water characteristics; second, it considers the unique spectral and phenological characteristics of tea gardens to highlight their differences from other evergreen vegetation; and third, it strengthens the distinguishing ability during key growth periods to improve the accuracy of remote sensing classification of tea gardens.

[0072] In the coarse classification stage, the research objects mainly involve land features such as construction land, water bodies, cultivated land, and evergreen vegetation. These categories have significant spectral differences, resulting in relatively small performance differences among the classifiers in the initial identification. Therefore, using three classifiers and selecting the best-performing result ensures the accuracy of the initial classification results, thus providing reliable prior information for subsequent fine classification. In the fine classification stage, this invention addresses the identification challenge caused by the high spectral similarity between tea gardens and other evergreen vegetation by employing a multi-model ensemble strategy. It fuses the probability prediction results of multiple classifiers through a soft voting mechanism, fully utilizing the complementary advantages of different method models to reduce the uncertainty of a single algorithm, thereby significantly improving the overall classification accuracy.

[0073] Furthermore, in the fine-grained classification stage, an ensemble learning method based on a soft voting mechanism is employed. First, based on the coarse classification, the evergreen vegetation category is further subdivided into tea gardens and other evergreen vegetation. This method uses the time-series features of the tea garden separation index during key phenological periods, combined with topographic parameters (elevation, slope, and aspect) to construct a feature space, integrating three machine learning algorithms. The soft voting mechanism is an advanced ensemble strategy, unlike traditional hard voting (majority rule), it fully utilizes the probabilistic prediction information of each classifier. For each sample, each classifier outputs a probability distribution; the soft voting mechanism obtains the final classification result by weighted averaging of these probability distributions, thus achieving higher accuracy.

[0074] To address potential salt-and-pepper noise and fragmented pixels after classification, this invention employs morphological processing techniques, including opening / closing operations and region connectivity analysis, to remove isolated misclassified pixels and enhance the spatial continuity and visual appeal of the results. In the post-processing stage, special measures were taken to address class imbalance, ensuring the effective identification of small evergreen vegetation patches. The resulting evergreen vegetation distribution map not only boasts high classification accuracy but also maintains good spatial integrity, laying a solid foundation for subsequent fine-grained differentiation between tea gardens and other evergreen vegetation.

[0075] The proposed "radar timing, optical classification" collaborative framework theoretically achieves functional complementarity of multi-source remote sensing data. Unlike traditional simple data overlay or feature-level fusion, this invention divides tasks in the spatiotemporal dimensions. Radar data is responsible for defining phenological periods in the temporal dimension, while optical data is responsible for classifying categories in the spatial dimension. The core advantage of this design lies in fully leveraging the inherent characteristics of different data sources: the all-weather observation capability of SAR data ensures accurate capture of key growth periods, while the rich spectral information of optical data provides a foundation for fine classification. This division of labor avoids the information redundancy and noise accumulation problems commonly found in multi-source data fusion. The application of radar data in the phenological analysis stage essentially utilizes its sensitivity to changes in vegetation structure to construct temporal constraints, providing an optimal time window for subsequent feature extraction from optical data. The introduction of this temporal constraint significantly improves the discriminative power of optical indices in tea garden identification, verifying the effectiveness of multi-source data collaboration.

[0076] Furthermore, this invention intentionally chose an ensemble learning model based on traditional machine learning rather than deep learning methods. This is because while the latter possesses powerful feature extraction capabilities, it is computationally intensive and resource-intensive, making it difficult to support efficient classification and real-time monitoring of large-scale tea gardens. In practical applications, especially in areas with limited computing resources, algorithms with high efficiency and low resource consumption are more valuable for widespread adoption. Therefore, this invention achieves an effective balance between the sophistication of the classification method and its feasibility under limited infrastructure conditions.

[0077] This invention, by constructing the TPRI index and establishing a multi-source data collaborative method for tea garden identification, not only achieves innovation in technical methods but also has significant practical value in the field of agricultural and forestry remote sensing applications. This method provides a new approach to solving the technical challenge of precise identification of economic crops and has positive significance for promoting the in-depth application of remote sensing technology in agriculture and forestry. From an industrial application perspective, accurate tea garden distribution information plays a crucial supporting role in tea industry planning, market analysis, and insurance assessment. This method can provide reliable basic data for relevant government departments and enterprises, supporting industrial policy formulation and business decisions. Furthermore, the crop identification approach based on phenological characteristics explored in this invention provides a new methodological framework for crop remote sensing monitoring, possessing strong theoretical innovation and methodological promotion value. From an engineering application perspective, future development should focus on the standardized deployment and industrial application of the algorithm. Exploring the direct deployment of the entire algorithm process to a cloud computing platform, utilizing distributed computing resources to process large-scale remote sensing data, and solving the problem of computing resource limitations is also important. Cloud platform deployment not only overcomes the limitations of single-machine computing power but also enables standardized algorithm services, providing unified tea garden identification capabilities for different users and supporting dynamic monitoring of tea gardens at regional, national, and even global scales. Simultaneously, this invention helps establish a comprehensive technical standard system and operational mechanism, including standardized data preprocessing procedures, parameter setting specifications, and accuracy evaluation standards. It also facilitates the development of user-friendly software tools and data service interfaces, promoting the widespread application and industrialization of the technology. Attached Figure Description

[0078] Figure 1 This is a flowchart illustrating a tea garden identification method based on multi-source remote sensing data according to a specific embodiment of the present invention.

[0079] Figure 2 This is a schematic diagram of a tea garden identification method based on multi-source remote sensing data according to a specific embodiment of the present invention.

[0080] Figure 3 This is a detailed flowchart of a tea garden identification method based on multi-source remote sensing data according to a specific embodiment of the present invention. Detailed Implementation

[0081] The present invention will now be described in detail with reference to the accompanying drawings.

[0082] The following detailed exemplary embodiments are disclosed. However, the specific structural and functional details disclosed herein are merely for the purpose of describing exemplary embodiments.

[0083] However, it should be understood that the present invention is not limited to the specific exemplary embodiments disclosed, but covers all modifications, equivalents, and substitutions falling within the scope of this disclosure. Throughout the description of the drawings, the same reference numerals denote the same elements.

[0084] Referring to the accompanying drawings, the structures, proportions, sizes, etc., depicted in the drawings are merely for illustrative purposes to aid those skilled in the art in understanding and reading the content disclosed herein. They are not intended to limit the conditions under which the invention can be implemented and therefore have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by the invention, should still fall within the scope of the technical content disclosed herein. Furthermore, the positional limitations used in this specification are merely for clarity of description and are not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention's implementation.

[0085] It should also be understood that the term “and / or” as used herein includes any and all combinations of one or more of the related listed items. Furthermore, it should be understood that when a component or unit is referred to as “connected” or “coupled” to another component or unit, it may be directly connected or coupled to the other component or unit, or there may be intermediate components or units. In addition, other words used to describe the relationship between components or units should be understood in the same manner (e.g., “between” versus “directly between,” “adjacent” versus “directly adjacent,” etc.).

[0086] Figure 1 This is a flowchart illustrating a tea garden identification method based on multi-source remote sensing data according to a specific embodiment of the present invention. As shown in the figure, the specific embodiment of the present invention includes a tea garden identification method based on multi-source remote sensing data, which includes the following steps:

[0087] S1: Acquire temporal radar data and optical remote sensing data of the target area;

[0088] S2: Extract key growth periods of tea gardens based on radar data, including the beginning and end of the growing season;

[0089] S3: Calculate vegetation index (VI) and water index (WI) using optical data, and select the VI and WI combination with the best discrimination by time-weighted dynamic time warping method.

[0090] S4: Determine the tea garden identification index as follows:

[0091] TRPI is the tea garden identification index, Ω1 is the linear operation, Ω2 is the nonlinear operation, T represents the tea garden sample, G represents the evergreen vegetation sample, and DTW is the dynamic time warping method.

[0092] S5: Tea garden identification is achieved based on the tea garden identification index and ensemble learning model.

[0093] This invention innovatively defines a novel tea garden identification index, TPRI, which combines vegetation and water indices to accurately map tea gardens. The fundamental reason TPRI can accurately identify tea gardens lies in its accurate capture of the essential characteristics of tea gardens as artificially managed ecosystems. Unlike natural evergreen forests, tea gardens experience frequent human disturbances throughout their growth cycle, particularly periodic harvesting activities, which significantly alter the canopy structure and water status, resulting in unique spectral response patterns. Traditional single vegetation indices can only reflect biomass changes and cannot comprehensively characterize this complex feature. By organically combining vegetation and water indices, TPRI not only reflects biomass changes in tea gardens but also simultaneously captures water status changes closely related to tea garden management, achieving a comprehensive expression of the multidimensional characteristics of the tea garden ecosystem. More importantly, this invention employs an index selection mechanism based on time-weighted dynamic time-regularized distance. By quantifying the temporal differences between tea gardens and other evergreen vegetation throughout their entire growth cycle, it ensures that the selected index combination has the strongest class distinguishing ability. This construction strategy, which maximizes the differences in phenological time series, enables TPRI to stably identify the unique growth rhythm characteristics of tea gardens under different geographical environments and climatic conditions, thereby achieving high-precision identification of tea gardens against a complex evergreen vegetation background.

[0094] The establishment of TPRI first investigated the key growth stages of tea gardens, and then designed an appropriate method combining vegetation and water indices to reflect the information differences between tea gardens and other evergreen vegetation to the greatest extent. The workflow of the specific implementation of this invention consists of the following two parts: (1) Derivation of key growth stages, including phenological analysis of tea gardens and other land cover types, and extraction of evergreen vegetation. (2) Determination of TPRI, its application in tea garden mapping, and verification of tea garden mapping results.

[0095] Furthermore, in the tea garden identification method based on multi-source remote sensing data in a specific embodiment of the present invention, the extraction of key growth stages of tea gardens based on radar data includes:

[0096] a) Perform radiometric calibration, geometric correction and filtering on radar data to construct a time series of VV / VH polarization backscattering coefficients;

[0097] b) For the VV / VH polarization backscattering coefficient time series, filter smoothing is performed. The formula for calculating the filter smoothing is:

[0098] ,

[0099] in, These are the values ​​after filtering and smoothing. These are the original values, i.e., the time series values ​​of the VV / VH polarization backscattering coefficients. The convolution coefficients are... The normalization coefficient is... is the half-width of the filtering window, and i is the offset relative to the center j of the sliding window, with a value ranging from -m to m;

[0100] This method achieves data smoothing through local polynomial fitting, offering the advantage of better preserving signal details while reducing noise compared to traditional mean or median filtering. The technical benefits of this smoothing process are threefold: First, it effectively suppresses the interference of speckle noise on the time series, making the phenological curves smoother and more continuous, facilitating subsequent derivative analysis and feature point extraction. Second, it maintains the temporal location and amplitude of key phenological events such as the beginning and end of the growing season, ensuring the accuracy of phenological parameter extraction. Finally, through reasonable setting of the window half-width parameter, an optimal balance is achieved between noise suppression and detail preservation, ensuring that the smoothed time series reflects the seasonal patterns of tea garden growth without excessive smoothing that could lead to the loss of key phenological features.

[0101] In a specific embodiment of the present invention, a dynamic threshold can also be calculated based on the smoothed scattering coefficient time series to filter valid observation data and remove outliers:

[0102] Threshold = min + 0.4 * (max - min)

[0103] in, Let be the threshold, and min and max be the minimum and maximum values ​​of the time series, respectively. This dynamic thresholding method has better adaptability than a fixed threshold, and can automatically adjust according to the growth characteristics of tea gardens in different regions. Key time nodes are identified by analyzing the changes in the first derivative of the phenological curves.

[0104] ,

[0105] The first derivative reflects the rate of change in the growth stage, and its extreme points correspond to key turning points in the vegetation growth state.

[0106] The start-of-season SOS and end-of-season EOS are determined based on the extreme points of the first derivative, as follows:

[0107] ,

[0108] ,

[0109] The argmax function is used to find the maximum value of a function within a given range. This is the change in the backscattering coefficient. This represents the corresponding time interval.

[0110] Accurate extraction of key growth stages in tea gardens is a crucial prerequisite for precise tea garden identification. Considering the limitations of optical remote sensing data, which is easily affected by weather conditions in cloudy and rainy areas, this invention proposes a method for identifying key phenological stages in tea gardens based on SAR radar data. SAR data has all-weather, all-time observation capabilities, and its backscattering coefficient is sensitive to changes in vegetation structure, effectively capturing structural feature changes in tea gardens during key stages such as growth and harvesting. As an artificially cultivated economic crop, tea gardens exhibit distinct seasonal characteristics in their growth cycle, mainly including the start of the growing season (SOS) and the end of the growing season (EOS). SOS typically corresponds to the spring bud sprouting and rapid growth stage, while EOS corresponds to the autumn growth slowdown and dormancy stage.

[0111] In this specific embodiment of the invention, Sentinel-1 SAR data is first radiometrically calibrated, geometrically corrected, and filtered to construct backscattering coefficient time series with VV and VH polarizations. VV polarization mainly reflects the vertical structural characteristics of vegetation, while VH polarization is more sensitive to vegetation scattering. The combination of the two polarization modes can comprehensively describe the structural and growth status changes of tea gardens. Subsequently, the Savitzky-Golay (SG) filtering algorithm is used to smooth the backscattering coefficient time series to eliminate noise and highlight phenological change trends. SG filtering achieves data smoothing through local polynomial fitting, and its mathematical expression is:

[0112] in, The smoothed value. This is the original data. The convolution coefficients are... The normalization coefficient is... is the half-width of the window, and i is the offset relative to the center j of the sliding window, with a value ranging from -m to m.

[0113] A dynamic threshold is calculated based on the smoothed time series data to filter valid observations and remove outliers.

[0114] ,

[0115] in, For the threshold, and These represent the minimum and maximum values ​​of the time series, respectively. A threshold value corresponding to a coefficient of 0.4 effectively distinguishes between normal and abnormal growth states in tea gardens: observations below this threshold typically correspond to signal attenuation due to cloud and rain, terrain shading effects, or data quality issues, while observations above this threshold represent the true growth state of the tea garden. In this specific embodiment of the invention, this coefficient was selected through comparative analysis, enabling more efficient differentiation of tea garden conditions. This dynamic threshold method has better adaptability than a fixed threshold, automatically adjusting according to the growth characteristics of tea gardens in different regions. Key time nodes are identified by analyzing the first derivative variation characteristics of the phenological curve:

[0116] ,

[0117] in, The backscattering coefficient at time t is... The first derivative reflects the rate of change in the growth stage, and its extreme points correspond to key turning points in the vegetation growth state.

[0118] The start of the growing season (SOS) is defined as the point where the first derivative reaches its maximum value during the period from the beginning to the middle of the year.

[0119] .

[0120] The end of the growing season (EOS) is defined as the point where the first derivative is minimized during the period from the middle of the year to the end of the year.

[0121] ,

[0122] in, The function is used to find the maximum value of a function within a given range. This is the change in the backscattering coefficient. This represents the corresponding time interval.

[0123] Furthermore, in the tea garden identification method based on multi-source remote sensing data in a specific embodiment of the present invention, the selection of the VI and WI combination with optimal discrimination through a time-weighted dynamic time warping method includes:

[0124] Constructing the Tea Garden Timeline With evergreen vegetation time sequence The weighted distance matrix is:

[0125] ,

[0126] in Point and The weighted distance between them; where, It can be given by prior knowledge or an empirical function, which measures the weighting of time deviation (or growth period difference, etc.) on the matching error at the i-th and j-th observation times.

[0127] Calculate the elements of the cumulative distance matrix:

[0128] ,

[0129] Construct the cumulative distance matrix ,in This represents the distance from the start of the sequence to the current position. The minimum cumulative distance;

[0130] The output time-weighted dynamic time warp distance is: .

[0131] The distance is TWDTW. The smaller the distance, the higher the similarity between the time series of the tea garden and other evergreen vegetation. Conversely, a larger distance indicates that there are significant differences between the two types of samples in their time series trajectories.

[0132] In addition, in the tea garden identification method based on multi-source remote sensing data of the present invention, the vegetation index includes at least one of the normalized vegetation index NDVI, enhanced vegetation index EVI, and ratio vegetation index RVI, and the water index includes at least one of the terrestrial water index LSWI, normalized water index NDWI, and modified normalized difference water index mNDWI.

[0133] In addition, the tea garden identification method based on multi-source remote sensing data of the present invention also includes a pre-extraction step of evergreen vegetation:

[0134] a) For each of the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Ratio Vegetation Index (RVI), Land Water Index (LSWI), Normalized Difference Water Index (NDWI), and Modified Normalized Difference Water Index (mNDWI), four time-series statistical features are extracted: annual mean, maximum, minimum, and annual variation. A 24-dimensional multi-temporal feature space is constructed. This time-series compression strategy condenses high-dimensional time series into low-dimensional statistics, which not only preserves key phenological information but also significantly reduces data dimensionality and computational complexity, avoiding the curse of dimensionality.

[0135] b) Combine random forest, support vector machine and extreme gradient boosting methods to classify land cover in the target area and extract evergreen vegetation areas;

[0136] c) Morphological processing methods, including opening and closing operations and region connectivity analysis, are used to remove isolated misclassified pixels and enhance the spatial continuity of the results.

[0137] In a specific embodiment of this invention, six key vegetation indices are calculated from time-series data: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Ratio Vegetation Index (RVI), Land Water Index (LSWI), Normalized Difference Water Index (NDWI), and Modified Normalized Difference Water Index (mNDWI). For each index, four temporal statistical features are extracted: annual mean, maximum value, minimum value, and annual variation (maximum value - minimum value), constructing a 24-dimensional multi-temporal feature space. This multi-index, multi-feature combination design can comprehensively capture the differences in phenological and spectral characteristics between evergreen vegetation and other land cover categories.

[0138] Subsequently, using this feature space data, three machine learning classifiers—Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost)—were used to classify land cover in the study area. To ensure the reliability of the classification results, a stratified random sampling strategy was adopted to generate training samples, while independent validation samples were set up to evaluate the classification performance. The performance of the three classifiers was systematically compared using comprehensive evaluation indices such as overall accuracy (OA), F1 score, and Kappa coefficient, as well as detailed confusion matrix analysis, and the best-performing classification result was selected. Among these, Random Forest, based on decision tree ensemble, excels at handling high-dimensional features and nonlinear relationships, and has strong robustness to noise and outliers; Support Vector Machine achieves complex nonlinear classification boundaries through kernel function mapping, maintaining good generalization ability even with small sample conditions; Extreme Gradient Boosting optimizes the loss function through the gradient boosting framework, automatically selecting features and handling class imbalance problems.

[0139] In one specific embodiment of the present invention, morphological processing techniques, including opening / closing operations and region connectivity analysis, are applied to address potential salt-and-pepper noise and fragmented pixels that may appear after classification. These techniques remove isolated misclassified pixels and enhance the spatial continuity and visual appeal of the results. In the post-processing stage, special measures are taken to address class imbalance, ensuring the effective identification of small evergreen vegetation patches. The resulting evergreen vegetation distribution map not only has high classification accuracy but also maintains good spatial integrity, laying the foundation for subsequent fine-grained differentiation between tea gardens and other evergreen vegetation.

[0140] Furthermore, in the tea garden identification method based on multi-source remote sensing data of the present invention, the tea garden identification is achieved using an ensemble learning model, including:

[0141] Integrating the probability outputs of three classifiers: Random Forest, Support Vector Machine, and Extreme Gradient Boosting:

[0142] ,

[0143] in, It is the final classification probability. is the probability prediction of the i-th classifier, and M is the number of classifiers;

[0144] A feature space is constructed by combining topographic parameters such as elevation, slope, and aspect.

[0145] In addition, in the tea garden identification method based on multi-source remote sensing data of the present invention, the radar data is Sentinel-1 C-band synthetic aperture radar SAR data, and the optical data is Sentinel-2 multispectral data.

[0146] For example, in one specific embodiment, this invention uses Sentinel-1 radar data and Sentinel-2 optical data as the basic data sources. For Sentinel-2 optical data, firstly, throughput filtering and cloud removal are performed using the QA band (Quality Assessment Band). To ensure spatial consistency across multiple time phases, precise geometric correction can be performed first, and the data from multiple periods can be constructed into a time-series dataset in chronological order. For Sentinel-1 SAR data, radiometric calibration is required to convert the original DN values ​​(DN values, Digital Numbers, are the original brightness values ​​of each pixel in the remote sensing image) into backscattering coefficients. Simultaneously, terrain correction is performed to eliminate the impact of terrain undulations on the radar signal. To improve data quality, appropriate filtering methods are also needed to reduce speckle noise in the SAR image.

[0147] To identify the most discriminative TPRI index, it is necessary to calculate the time series similarity of different remote sensing indices during the critical growth period. This invention first calculates the time series of three vegetation indices (EVI, RVI, NDVI) and three water indices (NDWI, mNDWI, LSWI), and standardizes each index to avoid the influence of orders-of-magnitude differences between indices on TWDTW distance comparisons. Subsequently, standardized index time series of tea garden samples and other evergreen vegetation are extracted. This invention employs a time-weighted dynamic time warping (TWDTW) algorithm to quantify the temporal differences between the two types of samples. This method is an improvement on the traditional DTW method; by introducing a time weighting function, it can better handle the seasonal variations of agricultural phenological characteristics. This method can effectively identify vegetation types with similar spectral characteristics but different phenological rhythms. Through statistical analysis, the vegetation index (VI) and water index (WI) with the largest TWDTW distance are selected, ensuring that the selected index combination has optimal discriminative power in tea garden identification, laying the foundation for subsequent TPRI index construction.

[0148] Tea gardens in the study area were classified using Sentinel-2 optical data with cloud cover <10%. In one specific embodiment of the invention, the first 10 bands of the Sentinel-2 optical data (bands 2-8, 8A, and 11-12) were used because they were designed for vegetation monitoring. To preserve the detailed spatial and spectral information provided by the 10m bands, the original bands with a spatial resolution of 20m were resampled to a spatial resolution of 10m. Therefore, a total of 10 bands of S2 optical data with a spatial resolution of 10m were used for tea garden identification.

[0149] Figure 2 This is a schematic diagram of a tea garden identification method based on multi-source remote sensing data according to a specific embodiment of the present invention. As shown in the figure, the specific embodiment of the present invention also includes a tea garden identification system based on multi-source remote sensing data, which includes a data acquisition module, a phenological analysis module, an index construction module, a classification module, and an output module.

[0150] The data acquisition module is used to acquire time-series radar data and optical remote sensing data of the target area;

[0151] The phenology analysis module is used to extract key growth stages in tea gardens based on radar data;

[0152] The index construction module is used to calculate the vegetation index (VI) and water index (WI) using optical data, and to select the optimal combination of VI and WI using a time-weighted dynamic time warping method; and to determine the tea garden identification index as follows:

[0153] Where Ω1 is a linear operation, Ω2 is a nonlinear operation, T represents the tea garden sample, G represents the evergreen vegetation sample, and DTW is the dynamic time warping method.

[0154] The classification module identifies tea gardens based on a tea garden identification index and an ensemble learning model.

[0155] The output module is used to generate a spatial distribution map of tea gardens based on the tea garden identification results.

[0156] This invention constructs the TPRI index by selecting the index with the largest TWDTW distance from both the VI and WI indices, fully utilizing the complementary advantages of vegetation indices reflecting biomass characteristics and water indices reflecting water conditions. Through both linear and nonlinear combination methods, the distinguishing ability of different index combinations is systematically evaluated, and ultimately, the combination that best distinguishes tea gardens from evergreen forests is selected to construct the TPRI index.

[0157] After construction, this invention verifies its effectiveness by comparing the TWDTW distance between TPRI and each individual index during the critical growth period. The verification logic is that if the TWDTW distance of TPRI exceeds that of all individual indices, it proves that the combination process indeed enhances the distinguishing ability of the original indices and achieves the expected goal of index-data construction; otherwise, it indicates that the combination method needs further optimization. This verification method ensures the consistency of evaluation criteria and the reliability of the combination effect.

[0158] Furthermore, in the tea garden identification system based on multi-source remote sensing data of the present invention, the phenological analysis module extracts the key growth stages of the tea garden based on radar data, including:

[0159] a) Perform radiometric calibration, geometric correction and filtering on radar data to construct a time series of VV / VH polarization backscattering coefficients;

[0160] b) For the VV / VH polarization backscattering coefficient time series, filter smoothing is performed. The formula for calculating the filter smoothing is:

[0161] ,

[0162] in, These are the values ​​after filtering and smoothing. This represents the original values, i.e., the time series of VV / VH polarization backscattering coefficients. The convolution coefficients are... The normalization coefficient is... is the half-width of the filtering window, and i is the offset relative to the center j of the sliding window, with a value ranging from -m to m;

[0163] The start-of-season SOS and end-of-season EOS are determined based on the extreme points of the first derivative, as follows:

[0164] ,

[0165] ,

[0166] The argmax function is used to find the maximum value of a function within a given range. This is the change in the backscattering coefficient. The corresponding time intervals are as follows. The growth season begins in the first half of the year (0-183 days), while the end of the growth season (EOS) occurs in the second half of the year (184-365 days). It should be noted that the days here are not necessarily synchronized with the dates of the calendar year.

[0167] Furthermore, in the tea garden identification system based on multi-source remote sensing data of the present invention, the index construction module is used to calculate the vegetation index VI and the water body index WI using optical data, and to screen the VI and WI combination with the best discrimination through a time-weighted dynamic time warping method, including:

[0168] Constructing the Tea Garden Timeline With evergreen vegetation time sequence The weighted distance matrix is:

[0169] ,in Point and Weighted distance between them;

[0170] Calculate the elements of the cumulative distance matrix:

[0171] ,

[0172] Construct the cumulative distance matrix ,in This represents the distance from the start of the sequence to the current position. The minimum cumulative distance;

[0173] The output time-weighted dynamic time warp distance is: .

[0174] Figure 3 This is a detailed flowchart of a tea garden identification method based on multi-source remote sensing data according to a specific embodiment of the present invention. The following is in conjunction with... Figure 3 The specific embodiments of the present invention will be described in more detail below.

[0175] As shown in the figure, the specific implementation of this invention includes the preprocessing of remote sensing data. In this specific implementation, Sentinel-1 SAR data is first subjected to data filtering, radiometric correction, geometric correction, and filtering. On the other hand, Sentinel-2 (S2) MSI images with cloud cover <10% are used to classify tea gardens in the area to be analyzed. The S2 images can be collected, for example, from Google Earth Engine (GEE), covering all areas to be analyzed. In this specific implementation, the first 10 bands (bands 2-8, 8A, and bands 11-12) are used because they are designed for vegetation monitoring. To preserve the detailed spatial and spectral information provided by the 10m bands, the original bands with a spatial resolution of 20m are resampled to a 10m spatial resolution. That is, a total of 10 S2 bands with a spatial resolution of 10m are used for tea garden mapping.

[0176] For the S2 image, data filtering, cloud and shadow removal, geometric correction, and resampling operations are performed based on its data characteristics, and vegetation index and water index are calculated accordingly.

[0177] Specifically, for Sentinel-1 SAR data, backscattering coefficient time series with VV and VH polarizations were constructed. VV polarization mainly reflects the vertical structure characteristics of vegetation, while VH polarization is more sensitive to vegetation scattering. The combination of the two polarization modes can comprehensively describe the structural and growth status changes of the tea garden. Subsequently, the SG filtering algorithm was used to smooth the backscattering coefficient time series to eliminate noise and highlight phenological change trends.

[0178] Following the SG filtering operation, a dynamic threshold calculation can be performed. The basic approach is to calculate the dynamic threshold based on the smoothed time series data, which is used to filter valid observation data and remove outliers. Based on the results of the dynamic prediction calculation, the start of the growing season (SOS) and the end of the growing season (EOS) can be determined.

[0179] Next, we proceed with the core step in this specific embodiment of the invention—the construction of the tea garden identification index—by extracting remote sensing information indices from Sentinel-2 (S2) MSI images. Specifically, this involves calculating six key vegetation or water body indices from time-series data: NDVI, EVI, RVI, LSWI, NDWI, and mNDWI. For each index, four temporal statistical features are extracted: annual mean, maximum value, minimum value, and annual amplitude (maximum value - minimum value), constructing a 24-dimensional multi-temporal feature space. Within this multi-temporal feature space, statistical features such as annual mean, maximum value, minimum value, and annual amplitude can be extracted from these indices.

[0180] In this specific embodiment of the invention, the time series of vegetation indices and water indices are standardized to avoid the influence of the order-of-magnitude differences between different indices on the TWDTW distance comparison. Subsequently, standardized index time series of tea garden samples and other evergreen vegetation are extracted separately. This invention uses TWDTW to quantify the temporal differences between the two types of samples. Statistical analysis is used to select the vegetation index (VI) and water index (WI) with the largest TWDTW distance, ensuring that the selected index combination has optimal discriminative ability in tea garden identification.

[0181] Then, the parameter space is exhausted, and different combination sequences are constructed by using linear or nonlinear combinations of VI and WI. Based on this, the tea garden identification index is derived.

[0182] Furthermore, in a specific embodiment of the present invention, the time-series curve of the tea garden identification index can be used in conjunction with RF, SVM, and XGBoost methods to classify land cover in the area to be analyzed. A soft voting mechanism is used to obtain the spatial classification results of the tea gardens. To ensure the reliability of the distribution results, a stratified random sampling strategy is used to generate training samples, while independent validation samples are set to evaluate the distribution effect. The performance of the three classifiers is systematically compared through comprehensive evaluation indices such as overall accuracy (OA), F1 score, and Kappa coefficient, as well as detailed confusion matrix analysis, and the classification result with the best performance is selected.

[0183] Therefore, the technical effects of this invention include the following.

[0184] The Time-Weighted Dynamic Time Warping (TWDTW) method not only considers the similarity of time series shapes but also incorporates time dimension matching constraints into the calculation, making the technical solution of this invention more sensitive to seasonal changes. This characteristic is particularly important when distinguishing vegetation types with similar spectral characteristics but different phenological rhythms (such as tea gardens and other evergreen vegetation). TWDTW more accurately quantifies the distance between time series by finding the optimal alignment path between two time series while penalizing excessive distortion on the time axis. First, time series of multiple vegetation indices and water indices need to be calculated. Then, the index time series of tea garden samples and other evergreen vegetation are extracted separately to ensure the integrity and comparability of the time series data. The time series distance between tea garden and other evergreen vegetation samples is calculated using TWDTW, and the vegetation index (VI) and water index (WI) with the largest TWDTW distance are selected through statistical analysis. These indices will be used for subsequent TPRI index construction.

[0185] The rationality of the TWDTW distance maximization principle is reflected on two levels: From a phenological perspective, the TWDTW distance can accurately measure the temporal differences between tea gardens and other evergreen vegetation throughout their growth cycle. A larger TWDTW distance means that the constructed tea garden identification index can better capture the unique growth patterns of tea gardens. From the perspective of index construction objectives, since the standard for selecting the optimal VI and WI is the maximum TWDTW distance, their combination (i.e., TPRI) should produce an even larger TWDTW distance. Otherwise, it indicates that this combination weakens rather than enhances the distinguishing ability of the original index, violating the original intention of index construction. Therefore, by verifying whether the TWDTW distance of the tea garden identification index exceeds that of all individual indices, the consistency of the evaluation criteria is ensured, and the combination process is guaranteed to truly play an optimization role.

[0186] The TPRI index determined in this specific embodiment of the invention is constructed based on the TWDTW distance maximization principle, a design with profound theoretical significance. The TWDTW algorithm, through the introduction of time weights, can more accurately quantify the temporal differences between land cover types with similar spectral characteristics but different phenological rhythms. In the specific application of tea garden identification, this method successfully captured the unique temporal response pattern of tea gardens during the harvesting period, a response pattern stemming from the periodic disturbance of the vegetation canopy structure by human management activities.

[0187] In terms of technical operability, the key steps involved in the construction process of the determined TPRI index, such as TWDTW calculation and SG filtering, all have mature open-source implementations, facilitating technology promotion and reproduction. In contrast, deep learning methods often require complex network architecture design, hyperparameter tuning, and large-scale GPU cluster support. Therefore, another advantage of the specific implementation of this invention lies in its strong interpretability; each processing step and parameter setting has a clear biophysical meaning, making it easy to understand and adjust, and improving adaptability in different application scenarios.

[0188] From an ecological perspective, tea gardens, as artificially managed agroforestry systems, exhibit phenological characteristics regulated by both the natural environment and human management practices. Frequent harvesting activities lead to canopy dynamics in tea gardens that differ significantly from natural evergreen vegetation during the growing season. This difference is the biological basis for the TPRI index's ability to effectively distinguish tea gardens from other evergreen vegetation. By organically combining vegetation indices with water indices, TPRI not only reflects changes in tea garden biomass but also captures changes in water conditions closely related to tea garden management, achieving a comprehensive expression of the multidimensional characteristics of the tea garden ecosystem.

[0189] This invention systematically explores and evaluates the discriminative power of different index combinations, ultimately selecting the index combination that best distinguishes tea gardens from evergreen forests, and thereby constructing the TPRI index specifically for identifying tea gardens. The choice to select one index from VI and one from WI, rather than simply choosing the two with the largest D_index, is based on the complementarity of these two types of indices in reflecting land cover characteristics: VI mainly reflects physiological characteristics such as vegetation biomass and photosynthesis, while WI focuses on reflecting the water content of vegetation and soil.

[0190] The Tea Garden Identification Index (TPRI) is constructed based on the following core ideas: First, it utilizes the complementarity of the selected VI and WI indices to comprehensively consider vegetation growth status and water characteristics; second, it considers the unique spectral and phenological characteristics of tea gardens to highlight their differences from other evergreen vegetation; and third, it strengthens the distinguishing ability during key growth periods to improve the accuracy of remote sensing classification of tea gardens.

[0191] In the coarse classification stage, the research objects mainly involve land features such as construction land, water bodies, cultivated land, and evergreen vegetation. These categories have significant spectral differences, resulting in relatively small performance differences among the classifiers in the initial identification. Therefore, using three classifiers and selecting the best-performing result ensures the accuracy of the initial classification results, thus providing reliable prior information for subsequent fine classification. In the fine classification stage, this invention addresses the identification challenge caused by the high spectral similarity between tea gardens and other evergreen vegetation by employing a multi-model ensemble strategy. It fuses the probability prediction results of multiple classifiers through a soft voting mechanism, fully utilizing the complementary advantages of different method models to reduce the uncertainty of a single algorithm, thereby significantly improving the overall classification accuracy.

[0192] Furthermore, in the fine-grained classification stage, an ensemble learning method based on a soft voting mechanism is employed. First, based on the coarse classification, the evergreen vegetation category is further subdivided into tea gardens and other evergreen vegetation. This method uses the time-series features of the tea garden separation index during key phenological periods, combined with topographic parameters (elevation, slope, and aspect) to construct a feature space, integrating three machine learning algorithms. The soft voting mechanism is an advanced ensemble strategy, unlike traditional hard voting (majority rule), it fully utilizes the probabilistic prediction information of each classifier. For each sample, each classifier outputs a probability distribution; the soft voting mechanism obtains the final classification result by weighted averaging of these probability distributions, thus achieving higher accuracy.

[0193] To address potential salt-and-pepper noise and fragmented pixels after classification, this invention employs morphological processing techniques, including opening / closing operations and region connectivity analysis, to remove isolated misclassified pixels and enhance the spatial continuity and visual appeal of the results. In the post-processing stage, special measures were taken to address class imbalance, ensuring the effective identification of small evergreen vegetation patches. The resulting evergreen vegetation distribution map not only boasts high classification accuracy but also maintains good spatial integrity, laying a solid foundation for subsequent fine-grained differentiation between tea gardens and other evergreen vegetation.

[0194] The proposed "radar timing, optical classification" collaborative framework theoretically achieves functional complementarity of multi-source remote sensing data. Unlike traditional simple data overlay or feature-level fusion, this invention divides tasks in the spatiotemporal dimensions. Radar data is responsible for defining phenological periods in the temporal dimension, while optical data is responsible for classifying categories in the spatial dimension. The core advantage of this design lies in fully leveraging the inherent characteristics of different data sources: the all-weather observation capability of SAR data ensures accurate capture of key growth periods, while the rich spectral information of optical data provides a foundation for fine classification. This division of labor avoids the information redundancy and noise accumulation problems commonly found in multi-source data fusion. The application of radar data in the phenological analysis stage essentially utilizes its sensitivity to changes in vegetation structure to construct temporal constraints, providing an optimal time window for subsequent feature extraction from optical data. The introduction of this temporal constraint significantly improves the discriminative power of optical indices in tea garden identification, verifying the effectiveness of multi-source data collaboration.

[0195] Furthermore, this invention intentionally chose an ensemble learning model based on traditional machine learning rather than deep learning methods. This is because while the latter possesses powerful feature extraction capabilities, it is computationally intensive and resource-intensive, making it difficult to support efficient classification and real-time monitoring of large-scale tea gardens. In practical applications, especially in areas with limited computing resources, algorithms with high efficiency and low resource consumption are more valuable for widespread adoption. Therefore, this invention achieves an effective balance between the sophistication of the classification method and its feasibility under limited infrastructure conditions.

[0196] This invention, by constructing the TPRI index and establishing a multi-source data collaborative method for tea garden identification, not only achieves innovation in technical methods but also has significant practical value in the field of agricultural and forestry remote sensing applications. This method provides a new approach to solving the technical challenge of precise identification of economic crops and has positive significance for promoting the in-depth application of remote sensing technology in agriculture and forestry. From an industrial application perspective, accurate tea garden distribution information plays a crucial supporting role in tea industry planning, market analysis, and insurance assessment. This method can provide reliable basic data for relevant government departments and enterprises, supporting industrial policy formulation and business decisions. Furthermore, the crop identification approach based on phenological characteristics explored in this invention provides a new methodological framework for crop remote sensing monitoring, possessing strong theoretical innovation and methodological promotion value. From an engineering application perspective, future development should focus on the standardized deployment and industrial application of the algorithm. Exploring the direct deployment of the entire algorithm process to a cloud computing platform, utilizing distributed computing resources to process large-scale remote sensing data, and solving the problem of computing resource limitations is also important. Cloud platform deployment not only overcomes the limitations of single-machine computing power but also enables standardized algorithm services, providing unified tea garden identification capabilities for different users and supporting dynamic monitoring of tea gardens at regional, national, and even global scales. Simultaneously, this invention helps establish a comprehensive technical standard system and operational mechanism, including standardized data preprocessing procedures, parameter setting specifications, and accuracy evaluation standards. It also facilitates the development of user-friendly software tools and data service interfaces, promoting the widespread application and industrialization of the technology.

[0197] The foregoing description illustrates and describes several preferred embodiments of the present invention. However, as mentioned above, it should be understood that the present invention is not limited to the forms disclosed in this specification and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described in this specification through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A tea garden identification method based on multi-source remote sensing data, characterized in that, The method comprises the steps of: S1: acquiring time-series radar data and optical remote sensing data of a target area; S2: extracting key growth periods of a tea garden based on the radar data, including a start period and an end period of the growth season; S3: calculating vegetation indices VI and water body indices WI using the optical data, and screening the optimal VI and WI combination with the highest discrimination degree by a time-weighted dynamic time warping method; S4: determining a tea garden identification index as follows: where Ω1is a linear operation, Ω2is a non-linear operation, T represents tea garden samples, G represents evergreen vegetation samples, and DTW is a dynamic time warping method. S5: realizing tea garden identification based on the tea garden identification index and an ensemble learning model; The screening of the optimal VI and WI combination with the highest discrimination degree by the time-weighted dynamic time warping method comprises: Constructing a tea garden chronology with evergreen vegetation chronologies weighted distance matrix is: wherein represents a weighted distance between points and ; calculating the elements of the cumulative distance matrix: ， Constructing a cumulative distance matrix wherein denotes the minimum cumulative distance from the start of the sequence to the current position ; The output time-weighted dynamic time warping distance is: .

2. The method for tea garden identification based on multi-source remote sensing data according to claim 1, characterized in that, extracting key growth periods of a tea garden based on the radar data comprises: a) performing radiation calibration, geometric correction and filtering processing on the radar data to construct a time series of VV / VH polarization backscattering coefficients; b) performing filtering smoothing on the time series of VV / VH polarization backscattering coefficients, and the calculation formula of the filtering smoothing is: ， wherein, is a filtered smoothed value, is an original value, is a convolution coefficient, is a normalization coefficient, is a filter window half-width, i is an offset relative to the center j of the sliding window, and the value range is -m to m. determining the start period SOS and the end period EOS of the growth season based on the extreme points of the first derivative, and the method is as follows: ， , Wherein, argmax function is used to find the maximum point of the function in the given range, is the change in the backscattering coefficient, is the corresponding time interval.

3. The method for tea garden identification based on multi-source remote sensing data according to claim 1, characterized in that, The vegetation indices include at least one of normalized difference vegetation index NDVI, enhanced vegetation index EVI and ratio vegetation index RVI, and the water body indices include at least one of land water body index LSWI, normalized difference water index NDWI and modified normalized difference water index mNDWI.

4. The method for tea garden identification based on multi-source remote sensing data according to claim 3, characterized in that, It also comprises a step of pre-extracting evergreen vegetation: a) for each of the normalized difference vegetation index NDVI, the enhanced vegetation index EVI, the ratio vegetation index RVI, the land water body index LSWI, the normalized difference water index NDWI and the modified normalized difference water index mNDWI, four time-series statistical features, i.e. annual mean, maximum value, minimum value and annual amplitude, are extracted to construct a 24-dimensional multi-temporal feature space; b) combining random forest, support vector machine and extreme gradient boosting method, performing feature classification on the target area to extract the evergreen vegetation area; c) for the classification result, a morphological processing method including opening and closing operation and region connectivity analysis is used to remove isolated misclassified pixels and enhance the spatial continuity of the result.

5. The method for tea garden identification based on multi-source remote sensing data according to claim 1, characterized in that, Realizing tea garden identification by using the ensemble learning model comprises: integrating the probability outputs of the three classifiers, i.e. random forest, support vector machine and extreme gradient boosting; ， wherein, is the final classification probability, is the probability prediction of the i-th classifier, and M is the number of classifiers. combining terrain parameters, i.e. elevation, slope and aspect, to construct a feature space.

6. The method for tea garden identification based on multi-source remote sensing data according to claim 1, characterized in that, The radar data is Sentinel-1 C-band synthetic aperture radar (SAR) data, and the optical data is Sentinel-2 multi-spectral data. 7.A tea garden identification system based on multi-source remote sensing data, characterized in that, It comprises a data acquisition module, a phenology analysis module, an index construction module, a classification module and an output module, wherein: the data acquisition module is used to acquire time-series radar data and optical remote sensing data of a target area; the phenology analysis module is used to extract key growth periods of a tea garden based on the radar data; the index construction module is used to calculate vegetation indices VI and water body indices WI using the optical data, and screen the optimal VI and WI combination with the highest discrimination degree by a time-weighted dynamic time warping method; and determine a tea garden identification index as follows: where Ω1is a linear operation, Ω2is a non-linear operation, T represents tea garden samples, G represents evergreen vegetation samples, and DTW is a dynamic time warping method. the classification module realizes tea garden identification based on the tea garden identification index and an ensemble learning model, The output module is configured to generate a tea garden spatial distribution map according to the tea garden identification result; The time-weighted dynamic time warping method is used to screen the combination of VI and WI with the optimal discrimination degree, and the combination of VI and WI with the optimal discrimination degree is determined as follows: Constructing a tea garden chronology with evergreen vegetation chronologies weighted distance matrix is: wherein represents a weighted distance between points and ; The elements of the cumulative distance matrix are calculated as follows: ， Constructing the cumulative distance matrix where denotes the minimum cumulative distance from the start of the sequence to the current position ; The output time-weighted dynamic time warping distance is: .

8. The tea garden identification system based on multi-source remote sensing data as claimed in claim 7, wherein, The phenology analysis module extracts key growth periods of the tea garden based on the radar data, and the method comprises the following steps: a) The radar data is radiometrically calibrated, geometrically corrected and filtered to construct a time series of VV / VH polarized backscatter coefficients; b) For the time series of VV / VH polarized backscatter coefficients, filtering and smoothing are performed, and the calculation formula of filtering and smoothing is as follows: ， wherein, is a filtered smoothed value, is an original value, is a convolution coefficient, is a normalization coefficient, is a filter window half-width, i is an offset relative to the center j of the sliding window, and the value range is -m to m. The beginning period (SOS) and the end period (EOS) of the growth season are determined based on the extreme points of the first derivative, and the method comprises the following steps: ， , Wherein, argmax function is used to find the maximum point of the function in the given range, is the change in the backscattering coefficient, is the corresponding time interval.

9. The tea garden identification system based on multi-source remote sensing data as claimed in claim 7, wherein, The exponential construction module is configured to calculate vegetation indices VI and water indices WI using optical data, and screen the combination of VI and WI with the optimal discrimination degree through a time-weighted dynamic time warping method, and the combination of VI and WI with the optimal discrimination degree is determined as follows: Constructing a tea garden chronology with evergreen vegetation chronologies weighted distance matrix is: wherein represents a weighted distance between points and ; The elements of the cumulative distance matrix are calculated as follows: ， Constructing a cumulative distance matrix wherein denotes the minimum cumulative distance from the start of the sequence to the current position ; The output time-weighted dynamic time warping distance is: .

Images